Concept-decision-knowledge post skill map construction method

Abstract

The invention discloses a concept-decision-knowledge post skill atlas construction method, which comprises the following steps of: constructing a post skill atlas model, and forming five hierarchical entity concepts; the knowledge decomposition unit is stored as a complete case or fragmentation rule; establishing a knowledge space, and explaining the space connotation of the post skill demand concept C, the decision information D and the knowledge resource K; and deep reinforcement learning is introduced, and a post skill knowledge graph construction scheme about demand identification, model construction and scheme conversion is provided. Through structural representation of the knowledge graph and perception reasoning of deep reinforcement learning, post skill knowledge utilization and task process recommendation are achieved, unified representation of a large amount of historical case knowledge and general rule knowledge of post occupational skills is achieved, and an exclusive domain model is provided for occupational skill identification.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and particularly relates to a method for constructing a job skill map of concept - decision - knowledge. Background Art

[0002] Vocational skill appraisal is a knowledge process regarding job tasks. Knowledge is the smallest element of the cognitive process and is often emphasized in cognitive reasoning - related work. Compared with machines, the greatest advantage of humans is the ability to use external information. Knowledge graphs are common media for machine - learning models to introduce external information. Introducing knowledge graphs in vocational skill appraisal can assist in analyzing and determining the optimal vocational skills and selecting accurate knowledge systems, which is an important process for clarifying the relationship between jobs and skill knowledge.

[0003] However, limited by the diversity of professional fields of job skills, job skill requirements are characterized by multitasks, multiple processes, multiple levels, multiple stages, multiple scenarios, multiple objects, etc. It mainly focuses on the practical learning of skills, and its knowledge representation method cannot yet maturely provide effective problem - solving solutions for decision - makers.

[0004] In recent years, many applications have achieved the representation of skill knowledge through methods such as graph convolutional networks, case - based reasoning, distributed planning, and cognitive diagnosis. However, these methods lack the connection between historical cases, uncertain problems, and new decision - making behaviors and do not pay attention to the impact of case knowledge reuse on job tasks and behaviors.

[0005] Therefore, the present invention proposes a method for constructing a job skill map of concept - decision - knowledge. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for constructing a job skill map of concept - decision - knowledge to solve the problems raised in the above - mentioned background art.

[0007] To achieve the above - mentioned purpose, the present invention provides the following technical solution: A method for constructing a job skill map of concept - decision - knowledge. S1: Describe the knowledge covering job types, job tasks, job skills, job characteristics, and job resource elements, and integrate knowledge models, knowledge units, and knowledge spaces;

[0008] S2: Construct a knowledge model, that is, construct a job skill map model;

[0009] S3: Decompose knowledge units and store them as complete cases or fragmented rules;

[0010] S4: Establish a knowledge space. Divide the generation of job skills and task - process solutions into three steps: determining initial requirements (concept C), selecting key information (decision D), and supplementing knowledge resources (knowledge K), and represent them as a CDK space using an adjacency matrix;

[0011] S5: Introduce deep reinforcement learning and propose a job skill knowledge construction plan for demand recognition, model construction, and solution transformation.

[0012] Preferably, in the above S2, the specific steps include:

[0013] S21: Use entities, relationships, and attributes of job vocational skills for knowledge representation of cases and rules, obtain a knowledge standardization structure, and clarify various elements such as the requirements, structure, and content of job vocational skills.

[0014] S22: Divide entity concepts into five levels, denoted as {P, R, S, F, T}; P represents the job type level with the label Post; R represents the job requirement level with the label Request; S represents the job skill level with the label Skill; F represents the skill feature level with the label Features; T represents the job task level with the label Task; each level represents the element category of job vocational skills and task processes, and each element consists of the corresponding label, name, and attributes.

[0015] S23: Define the relationships between entities in the job skill knowledge graph model.

[0016] Preferably, the specific steps of the above S3 include: S31: Establish a case knowledge unit, defined as a case knowledge unit triple containing Case_Head, Case_Relation, and Case_Tail, denoted as Case_Units = <Case_Head, Case_Relation, Case_Tail>.

[0017] S32: Establish a rule knowledge unit, defined as a rule knowledge unit triple containing Rule_Head, Rule_Relation, and Rule_Tail, denoted as Rule_Units = <Rule_Head, Rule_Relation, Rule_Tail>.

[0018] Preferably, the above S4 specifically includes:

[0019] S41: Create a C space;

[0020] S41-1: Create the domain connotation of the C space, represent the job skill requirement concept as C nc , and map various types of skills to the demand matrix C MIn; according to the knowledge unit representation, a space C that meets the specifications of the knowledge point unit in the vocational skill appraisal process is formed by using the node table N(C) and the adjacency matrix A(C). The space C covers job skill concepts at multiple levels;

[0021] S41-2: Create a representation of space C. Construct the node table N(C) to include instantiated and non-instantiated knowledge entities or nodes composed of labels and attributes; in the adjacency matrix A(C), the [0,1] value indicates whether there is a connection between the starting entity in the corresponding row and the ending entity in the corresponding column. 0 indicates no connection, and 1 indicates a connection:

[0022] S42: Create space D:

[0023] S42-1: Create the domain connotation of space D. Represent the decision-making space of job skills and task processes as a set of key concepts D that constitute the design solutions for skills and task processes nd ; According to the knowledge unit representation, space D includes the smallest subgraph in the knowledge unit that covers the job skill and task process design requirement entities; map the requirements and related concepts proposed by the decision-maker, which are the main factors for solution generation, to the decision matrix D M Including the node table N(D) and the adjacency matrix D), a decision-making space D is formed;

[0024] S42-2: Create a representation of space D. The size of space D is determined by the number of nodes in the smallest subgraph of the key entities. The [0,1] value in the adjacency matrix A 0 (D) indicates whether there is a connection between the starting entity in the corresponding row and the ending entity in the corresponding column. 0 indicates no connection, and 1 indicates a connection;

[0025] S43: Create space S:

[0026] S43-1: Create the domain connotation of space S. Represent the knowledge resource information as K nk and map the related concepts uniformly to the resource matrix K M As a branch element for solution generation, including the node table N(K) and the adjacency matrix A(K), a supplementary resource space K that meets the specifications of the job skill and task process knowledge unit is formed;

[0027] S43-2: Create a representation of space S. The node table N(K) includes instantiated and non-instantiated knowledge entities or nodes composed of labels and attributes. The [0,1] value in the adjacency matrix A(K) indicates whether there is a connection between the starting entity in the corresponding row and the ending entity in the corresponding column. 0 indicates no connection, and 1 indicates a connection.

[0028] Preferably, the said S5 specifically includes:

[0029] S51: Conduct requirement recognition; match the job skills and task processes with the nodes in the knowledge graph, and convert them into a feature adjacency matrix with a specific structure and connotation according to the CDK space representation, forming an initial requirement state space based on deep reinforcement learning;

[0030] S51-1: Process requirement representation. Using the knowledge units of job skills and task processes, the decision maker identifies them through word segmentation and named entity natural language processing technologies, and extracts relevant process concepts from the design requirement text to form a requirement list composed of labels and attributes;

[0031] S51-2: Solution representation. Based on the CDK knowledge representation definition of job skills and task processes, the instantiated representation of the initial process requirements is integrated into the process requirement adjacency matrix A 0 (C); meanwhile, the process of generating solutions for job skills and task processes is abstracted as adding instantiated nodes and relationships to the adjacency matrix. Therefore, the features in A 0 (C) are assigned to the D and K spaces;

[0032] S51-3: Assign the concepts in the C space to the D and K spaces. According to the knowledge unit representation, select the smallest subgraph that covers the proposed process requirement concepts; the concepts within the smallest subgraph are assigned to the D space to form an adjacency matrix A 0 (D) representing job skills and task process information; then, assign the concepts not in the smallest subgraph to the K space to generate an adjacency matrix A 0 (K);

[0033] S52: Conduct model construction. Design a deep deterministic policy gradient reinforcement learning algorithm. Through joint training of the actor network for learning the policy and the critic network for learning the state value, use a deep neural network as an approximation of the policy network and the action value function, and use stochastic gradient descent to train the parameters of the policy network and the value network model. Introduce a twin neural network to achieve a more stable learning process;

[0034] S52-1: Construct the state space. Represent the structure and content state of the solution as a state matrix S, and use the adjacency matrix and node table to represent the structure and content of the solution respectively;

[0035] S52 - 2: Construct the action space, update the structure and content of job skills and task processes using stored case or rule knowledge, and transform the process of adding knowledge units with the <head, relation, tail> triple structure to the state space into instance selection for the head entity or tail entity in the node table, which is used to create relationships at the non - diagonal positions of the adjacency matrix; represent each action as an action matrix A, forming a matrix with a dimension of A_dim = 2 * 2([Vector h , Index h ,[Vector t , Index t );

[0036] S53: Design a reward mechanism that includes regional weights, unit rewards, and result evaluation;

[0037] S53 - 1: Create regional weights, obtain the connotation of the job skills process by merging the D and K feature spaces, and divide the state space into three regions: D, DK, and K;

[0038] S53 - 2: Create unit rewards, divide the reward direction according to the newly added process knowledge after state update, and find the corresponding case or rule knowledge units for the <head, relation, tail> triple corresponding to the action space A t , denoted as A t ∈ {Case_Units, Rule_Units});

[0039] S53 - 3: Create result evaluation, calculate the cosine similarity SS t between the adjacency matrix A(S t ) of each step's state space S and the adjacency matrix A(PS_KGM) corresponding to the knowledge unit t .

[0040] Preferably, in S52 - 1, the initial state space S 0 of the input model is the initial solution space A(D&K) obtained from requirement identification; based on the initial state space, the solution generation process involves updating the state space; the instantiation of each node vector in the state matrix node table represents adding 1 job skill entity to the solution.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] Through the structured representation of the knowledge graph and the perceptual reasoning of deep reinforcement learning, the present invention solves the utilization of job skill knowledge and task process recommendation, realizes the unified representation of a large number of historical case knowledge and general rule knowledge of job vocational skills, and provides a domain - specific model for vocational skill appraisal. Brief Description of the Drawings

[0043] Figure 1 It is a schematic flowchart of the present invention;

[0044] Figure 2 It is a generated diagram of the technical solution of the present invention;

[0045] Figure 3 It is a diagram for extracting skill requirement nodes of the present invention;

[0046] Figure 4 It is a conversion diagram of the solution of the present invention;

[0047] Figure 5 It is a flowchart of the deep reinforcement learning algorithm of the present invention; Detailed Description of the Invention

[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the protection scope of the present invention.

[0049] Embodiment 1

[0050] Please refer to Figure 1 and Figure 5 , the present invention provides a technical solution: Figure 1 As shown in, a method for constructing a job skill map of concept - decision - knowledge of the present invention is implemented on an application platform. The platform includes three parts: a knowledge model, knowledge units, and a knowledge space.

[0051] Step S1: Describe the knowledge of elements such as job types, job tasks, job skills, job characteristics, and job resources, integrate the knowledge model, knowledge units, and knowledge space, and construct a "concept - decision - knowledge" job skill map framework.

[0052] Step S2: The knowledge model is a job skill knowledge map model, and a five - layer entity concept model is constructed, denoted as {P, R, S, F, T}. Define the relationships between entities in the job skill knowledge map model, including bind_to representing the binding relationship between job types and job requirements, relate_to representing the dependency relationship between job requirements and job skills, has_auxiliary representing the auxiliary relationship between job skills and skill characteristics, has_choice representing the selection relationship between skill characteristics and tasks, and has_contain representing the inclusion relationship between job requirements and tasks.

[0053] As an alternative, in the above-mentioned job skill knowledge graph, a job can have multiple requirements and skills, which are connected through hierarchical construction and relationships such as bind_to, relate_to, has_auxiliary, has_choice, has_contain, etc., respectively forming job requirement and skill sets {Post 1 , Post 2 ,......, Post n}{Request 1 , Request 2 ,...... Request n}{Skill 1 , Skill 2 ,......, Skill n}{Post 1 , Post 2 ,......, Post n}{Features 1 , Features 2 ,......, Features n}{Task 1 , Task 2 ,...... Task n}; Through the relationship connection, a job requirement can have multiple skills, a job skill has multiple features, and a feature can have multiple specific tasks. In this way, the knowledge graph model is gradually constructed from concept - attribute - instance.

[0054] Step S3: As Figure 1 shown, knowledge units are divided into two types: case knowledge units and rule knowledge units, and their structure is represented as a triple <head, relation, tail>. The case knowledge unit is represented as Case_Units = <Case_Head, Case_Relation, Case_Tail>, and the rule knowledge unit is represented as Rule_Units = <Rule_Head, Rule_Relation, Rule_Tail>.

[0055] Step S4: As Figure 1 shown, the construction of the job vocational skill knowledge space is divided into three steps: determining the initial requirements (concept C), selecting key information (decision D), and supplementing knowledge resources (knowledge K), and is represented as the CDK space using an adjacency matrix.

[0056] Step S41, create the C space

[0057] Step S41-1: Create the connotation of the C space domain. Before formulating the process solution for post skills, it is necessary to clarify specific post requirements, including concepts such as identifying the post skills to be identified, defining existing task resources, and the target skills or work tasks to be emphasized. Represent the design requirement concepts of the vocational skill appraisal process as C nc , including entity labels and attributes. Defined as:

[0058] C nc ={Label_C nc ,Attribute_C nc}

[0059] Label_C nc ,Attribute_C nc ∈MP_KGM

[0060] To promote the alignment of key information between design requirements and knowledge units, map various types of skills to a requirements matrix C M . According to the knowledge unit representation, use the node table N(C) and the adjacency matrix A(C) to form a space C that meets the specification requirements of the knowledge point units in the vocational skill appraisal process. C covers post skill concepts at multiple levels.

[0061] Step S41-2: Create the C space representation. The node table N(C) contains instantiated and non-instantiated knowledge entities or nodes composed of labels and attributes. The [0,1] value in the adjacency matrix A(C) indicates whether there is a connection between the starting entity in the corresponding row and the ending entity in the corresponding column. 0 indicates no connection, and 1 indicates a connection.

[0062] Step S42: Create the D space.

[0063] Step S42-1: Create the connotation of the D space domain. The decision space representation of post skills and task processes constitutes a set of key concepts D nd for the design solution of skills and task processes. According to the knowledge unit representation, the D space includes the smallest subgraph that covers the post skill and task process design requirement entities in the knowledge unit. Defined as:

[0064] D nd ={Label_D nd ,Attribute_D nd}

[0065] Label_D nd ,Attribute_D nd ∈MP_KGM

[0066] After describing the main job skill requirements, information related to tasks, skills, characteristics, resources, and other factors is given priority; then, the job skill concepts closely related to the above information are integrated into the D space of the task process as the key entities for unified representation. To simplify the matching and calculation of related concepts, the requirements and related concepts proposed by the decision maker are used as the main factors for solution generation and mapped to the decision matrix D M which includes the node table N(D) and the adjacency matrix D), forming a decision space D that conforms to the job skill and task process knowledge unit specifications.

[0067] Step S42-2: Create the D space representation. The size of the D space is determined by the number of nodes in the minimum subgraph of the key entity, i.e., nd 2 of the adjacency matrix A(D), where nd is the number N of entity label types in the minimum subgraph of the basic entity mingraph which is less than or equal to the total number N of entity label types in the top-level knowledge unit MP_KGM . The adjacency matrix A 0 (D) has [0,1] values indicating whether there is a connection between the starting entity in the corresponding row and the ending entity in the corresponding column. 0 indicates no connection, and 1 indicates a connection.

[0068] Step S43: Create the S space.

[0069] Step S43-1: Create the S space domain connotation. To generate a global job skill and task process that conforms to the constraints of the knowledge unit, it is necessary to add additional process information K nk beyond the initial requirements to assist the decision maker in process design. It is defined as:

[0070] K nk ={Label_K nk ,Attribute_K nk}

[0071] Label_K nk ,Attribute_K nk ∈MP_KGM

[0072] Unify the related concepts (concepts not explicitly stated by the technical personnel but specified by the knowledge unit) and map them to the resource matrix K M as the branch elements for solution generation, including the node table N(K) and the adjacency matrix A(K), forming a supplementary resource space K that conforms to the job skill and task process knowledge unit specifications and plays an auxiliary role in the solution generation process.

[0073] Step S43-2: Create S space representation. The node table N(K) includes instantiated and non-instantiated knowledge entities or nodes consisting of labels and attributes. The size of K space is determined by the number of nodes in the supplementary resources, i.e., nk 2 The adjacency matrix A(K) is nk, which is the total number of entity tag types N in the knowledge model. MP_KGM Subtract the number of entity tag types N in the smallest subgraph of the key entity mingraph The [0,1] value in the adjacency matrix A(K) indicates whether there is a connection between the start entity of the corresponding row and the end entity of the corresponding column, 0 indicates no connection, and 1 indicates a connection.

[0074] Step S5: Figure 2 As shown, generate the solution

[0075] Step S51: Perform demand identification. Match job skills and task processes with nodes in the knowledge graph, and convert them into feature adjacency matrices with specific structures and connotations based on the representation of the CDK space to form an initial demand state space based on deep reinforcement learning.

[0076] Step S51-1: Process requirement expression. Figure 3 As shown in the figure, using the knowledge units of job skills and task processes, decision makers use natural language processing techniques such as word segmentation and named entities to identify and extract relevant process concepts from the design requirement text to form a requirement list consisting of labels and attributes. The positions are derived from the standardized positions stipulated by the state, the "Classification of Occupations of the People's Republic of China (2022 Edition)" and the "National Teaching Standard System for Vocational Education" of the Vocational Education Department of the Ministry of Education. Extraction and text representation of process requirement nodes. To ensure that the extracted process nodes match the entities stored in the skill knowledge graph, the label Label:Name and attribute Attribute:Value information contained in the node are combined into a short text. The text2vec text representation tool is used to calculate the similarity between the process requirement node and the knowledge graph node, and the matching result with higher similarity is selected as the initial instantiation representation of the process requirement.

[0077] Step S51-2: Solution presentation. Figure 4 As shown in Figure 1, based on the job skills and task process CDK knowledge representation definition, the instantiation representation of the initial process requirements is integrated into the process requirement adjacency matrix A. 0 (C). At the same time, the process of generating job skills and task process solutions is abstracted to adding instantiated nodes and relationships to the adjacency matrix. 0 (C) The features are allocated to D and K spaces.

[0078] Step S51-3: To convert the requirements space into a solution space, the concepts in the C space are assigned to the D and K spaces. According to the knowledge unit representation, the smallest subgraph covering the concepts of the proposed process requirements is selected. The concepts within the smallest subgraph are assigned to the D space to form the adjacency matrix A representing the job skills and task process information 0 (D); then, the concepts not in the smallest subgraph are assigned to the K space to generate the adjacency matrix A 0 (K).

[0079] Step S51-3-1: Using the combination method, unordered pairs of requirement labels are selected and represented as Start C and End C .

[0080] Step S51-3-2: In the knowledge model, the undirected knowledge chains that meet the shortest length requirement are matched, and two requirement labels are selected as the start and end, represented as <Start C ,Label 1 ,...,Label x ,End C >.

[0081] Step S51-3-3: Combine the combined matches of all knowledge chains, remove the duplicate start, end, and intermediate labels, obtain the smallest knowledge subgraph covering the initial requirements, and represent it as the adjacency matrix A 0 (D), with the matrix dimension of p, and the adjacency list is arranged in the way that the pre-order is the instantiated label and the post-order is the non-instantiated label.

[0082] Step S51-3-4: Treat the remaining non-instantiated process labels as supplementary resources, represented as A 0 (K), with the matrix dimension of q.

[0083] Step S51-3-5: Combine A 0 (D) and A 0 (K) into the adjacency matrix A 0 (D&K), representing the initial solution space S 0 . The upper left p*p area of the matrix is the adjacency matrix of the D space, the lower right q*q area is the adjacency matrix of the K space, and the values of the remaining areas are defaulted to 0.

[0084] Step S51-3-6: The obtained solution space S 0 only contains the instantiated representation of the initial requirements, and the remaining entities and relationships await instantiation generation by the reinforcement learning model. Therefore, S 0 is provided as input to the reinforcement learning model to participate in model training.

[0085] Step S52: Build a model. In combination with the CDK model representation, configure states, actions, and rewards for deep reinforcement learning, build a generation process solution based on job skills, task process cases, and rule knowledge, and obtain the adjacency matrix of the solution.

[0086] Further, as Figure 5 shown, design a deep deterministic policy gradient reinforcement learning algorithm. By jointly training an actor network for learning policies and a critic network for learning state values, use a deep neural network as an approximation of the policy network and the action value function. Use stochastic gradient descent to train the parameters of the policy network and the value network model. Introduce a twin neural network to achieve a more stable learning process, enabling the job skill knowledge graph to gradually improve the policy and learn the continuous action space, and make better decision-making capabilities.

[0087] Step S52-1: Build the state space. The state space represents the structure and content of the job skill design solution. The structural form of the solution is defined by the knowledge units of job skills and task processes, representing conceptual elements (e.g., types, tasks, features, etc.) and their relationships (e.g., bindings, inclusions, involvements, etc.). During the solution generation process, reinforcement learning gradually improves the content of the solution by obtaining case or rule knowledge from the job skill knowledge graph. Represent the structural and content states of the solution as a state matrix S, and use the adjacency matrix and the node table to represent the structure and content of the solution respectively.

[0088] Further, the initial state space S of the input model 0 is the initial solution space A(D&K) obtained for requirement identification. Based on the initial state space, the solution generation process involves updating the state space. The instantiation of each node vector in the state matrix node table represents adding 1 job skill entity to the solution. On the contrary, each change from 0 to 1 in the non-diagonal positions of the adjacency matrix represents the creation of a new relationship between job skills and task processes. After the state space is updated, the matrix result S output by the model t+1 represents the job skill and task process solution.

[0089] Step S52-2: Build the action space. The action space task uses the stored case or rule knowledge to update the structure and content of job skills and task processes, and transforms the process of adding knowledge units with the <head,relation,tail> triple structure to the state space into instance selection for the head entity or the tail entity in the node table, for creating relationships in the non-diagonal positions of the adjacency matrix. Represent each action as an action matrix A, forming a dimension of A_dim = 2*2([Vector h ,Index h ,[Vectort , Index t ) matrix, which reduces the dimension of the action space without affecting the ability of the action matrix A to update the state matrix S.

[0090] Furthermore, the action space A t performs the following steps in each action:

[0091] Step S52-2-1: t = 0.

[0092] Step S52-2-2: Randomly select an instantiated entity from the initial state space S 0 as the head entity head, and record its index position in the node table as the row.

[0093] Step S52-2-3: Establish a new relationship. Starting from the starting position [row, row] of the initial state space S0, randomly change a 0 value to 1 at a non-starting point in the same row to create a relationship starting from the head and represent the new relationship index as [row, col].

[0094] Step S52-2-4: According to the relationship index, select the process entity at the index position col in the node table as the tail entity and determine whether the tail entity is instantiated. If it is not instantiated (value is 0), randomly match the entity vector corresponding to the case or rule knowledge from the post skills and task process knowledge graph; otherwise, keep the original instantiated vector representation of the post skills knowledge.

[0095] Step S52-2-5: Through the above steps, obtain a relationship triple with both the head entity and the tail entity instantiated; then, deduce the action matrix A of the elements' indexes in the node table, adjacency matrix, and triple t . Calculate S t ∪A t , and use the values in A t to replace the values at different positions to obtain the updated state space S t+1 .

[0096] Step S52-2-6: Store each executed action matrix A t in the action space A m to check for action repeatability.

[0097] Step S52-2-7: When , establish a new relationship based on the action space A m in step ②, and do not repeat establishing a relationship at the position where a relationship has already been established.

[0098] Step S52-2-8: t = t + 1, and loop through the above steps until the termination condition is met.

[0099] Step S53: Design a reward mechanism that includes regional weights, unit rewards, and result evaluation.

[0100] Step S53-1: Create regional weights. Obtain the connotations of job skills and task processes by merging the D and K feature spaces, and divide the state space into three regions: D, DK, and K.

[0101] Step S53-1-1: The state update in the D region is based on the instantiated representation expansion of the proposed job skill requirements. To ensure that the solution is consistent with the requirements of technicians, emphasize the use of existing instance information and set a reward weight W for the D region. 1 。

[0102] Step S53-1-2: The state update in the DK region represents the creation of nodes and relationships between key entities and supplementary resources in job skills and task processes. Since the state of the D region is crucial for the solution, the creation of relationships in the DK region related to it should also be rewarded, and a reward weight W is set for the DK region. 2 。

[0103] Step S53-1-3: The state update in the K region represents the creation of nodes and relationships between supplementary resources, which are not closely related to the initial process design requirements but are necessary for a complete solution to job skills and task processes, and a reward weight W is set for the K region. 3 。

[0104] Furthermore, based on the above regional division, a weight relationship of W 1 >W 2 >W 3 is established. Let the initial reward be reward t =1, and each reward is calculated by weighted calculation as reward t =reward t *W i , i ∈ [1, 2, 3].

[0105] Step S53-2: Create unit rewards. Define the reward direction and value for matching case and rule unit knowledge. According to the newly added process knowledge after state update, divide the reward direction. If the newly added knowledge is stored in the current case and rule knowledge graph, that is, the triple <head, relation, tail> in the action space A t finds the corresponding case or rule knowledge unit, denoted as A t ∈ {Case_Units, Rule_Units}).

[0106] Furthermore, if the new knowledge does not exist in the knowledge graph, the reward direction is expressed as reward d=P j , where j∈[1,2,3]. 1 Indicates the reward direction when the new knowledge belongs to the case knowledge unit and is greater than 0; P 2 Indicates the reward direction when the new knowledge belongs to the rule knowledge unit and is greater than 0; P 3 Indicates the reward direction when the new knowledge does not belong to the existing knowledge and is less than 0. Since the case knowledge unit is embedded in the complete job skills and task process cases, there is a strong structural association between them. A job skills solution with case knowledge as the main component and rule knowledge as the supplement is established. j With P 1 >P 2 >P 3 relation.

[0107] Furthermore, when the average similarity is greater, the new knowledge is more consistent with job skill knowledge. 3 In this case, if the new knowledge has no corresponding knowledge unit, the reward value R 3 =P 3 Combining the area weight and reward direction, each reward value is defined as:

[0108]

[0109] Step S53-3: Create result evaluation. Calculate the state space S for each step t The adjacency matrix A(S t ) and the cosine similarity SS between the adjacency matrix A(MP_KGM) corresponding to the knowledge unit t . When the similarity between the state matrix and the knowledge unit increases, the reward value increases. SS t It is expressed as:

[0110] SS t =cos(A(S t ),A(MP_KGM))

[0111] SS t+1 =cos(A(S t+1 ),A(MP_KGM))

[0112]

[0113] Furthermore, by obtaining the reward value of each state update after executing the action, the reward t =reward t *factor, and accumulate the total reward value. When the model ends, the final state space S is output t+1, including a node table and an adjacency matrix. To facilitate the accumulation of job skill knowledge, the text vectors in the node table are parsed into entities in the knowledge graph. In addition, according to the knowledge unit representation, each relationship is assigned a label to obtain a knowledge graph, which allows decision-makers to retrieve detailed information about the generated solutions through the knowledge graph.

[0114] Although the embodiments of the present invention have been shown and described, see the above detailed description, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a concept-decision-knowledge job skill map, characterized by: S1: Describe the knowledge of job types, job tasks, job skills, job characteristics, and job resource elements, and integrate knowledge models, knowledge units, and knowledge spaces; S2: Build a knowledge model, that is, build a job skill map model; S3: Decompose knowledge units and store them as complete cases or fragmented rules; S4: Establish a knowledge space, divide the generation of job skills and task process solutions into three steps: determining initial requirements (concept C), selecting key information (decision D), and supplementing knowledge resources (knowledge K), and use the adjacency matrix to represent it as CDK space; S5: Introduce deep reinforcement learning and propose job skill knowledge construction plans for demand identification, model construction, and solution transformation.

2. The method for constructing a concept-decision-knowledge job skill map according to claim 1, characterized in that: In said S2, the specific steps include: S21: Through the entities, relationships, and attributes of job skills, the knowledge representation of cases and rules is used to obtain the standardized structure of knowledge and clarify the requirements, structure, and content of job skills; S22: The entity concept is divided into five levels, represented by {P, R, S, F, T}; P represents the job type layer, labeled Post; R represents the job requirement layer, labeled Request; S represents the job skill layer, labeled Skill; F represents the skill feature layer, labeled Features; T represents the job task layer, labeled Task; Each level represents the element category of job professional skills and task processes, and each element consists of corresponding labels, names and attributes; S23: Define the relationship between entities in the job skill knowledge graph model.

3. The method for constructing a concept-decision-knowledge job skill map according to claim 1, characterized in that: The specific steps of S3 include: S31: establishing a case knowledge unit, which is defined as a case knowledge unit triple containing Case_Head case head, Case_Relation case relationship and Case_Tail case tail, expressed as Case_Units=<Case_Head,Case_Relation,Case_Tail> ; S32: Establish a rule knowledge unit, which is defined as a rule knowledge unit triple containing Rule_Head, Rule_Relation and Rule_Tail, expressed as Rule_Units=<Rule_Head,Rule_Relation,Rule_Tail> .

4. The method for constructing a concept-decision-knowledge job skill map according to claim 1, characterized in that: The S4 specifically includes: S41: Create C space; S41-1: Create the C space domain connotation and express the concept of job skill requirements as C nc , mapping various types of skills to the demand matrix C M According to the knowledge unit representation, the node table N(C) and the adjacency matrix A(C) are used to form a space C that meets the requirements of the knowledge point unit specification of the vocational skill appraisal process, and C covers multiple levels of job skill concepts; S41-2: Create a C-space representation and construct a node table N(C) containing instantiated and non-instantiated knowledge entities or nodes consisting of labels and attributes; the [0,1] value in the adjacency matrix A(C) indicates whether there is a connection between the start entity of the corresponding row and the end entity of the corresponding column, 0 indicates no connection, and 1 indicates a connection: S42: Create D space: S42-1: Create D-space domain connotations and represent the decision space of job skills and task processes as a set of key concepts that constitute the design solution for skills and task processes. nd ; According to the knowledge unit representation, the D space includes the smallest subgraph covering the job skills and task process design requirement entities in the knowledge unit; the requirements and related concepts proposed by the decision maker are used as the main factors for solution generation and mapped to the decision matrix D M In, including the node table N(D) and the adjacency matrix D), a decision space D is formed; S42-2: Create a D space representation, where the size of the D space is determined by the number of nodes in the smallest subgraph of the key entity. The [0,1] value in the adjacency matrix A0(D) indicates whether there is a connection between the start entity of the corresponding row and the end entity of the corresponding column, 0 indicates no connection, and 1 indicates a connection; S43: Create S space: S43-1: Create the S space domain connotation and represent the knowledge resource information as K nk , map related concepts uniformly to the resource matrix K M On the top, as the branch elements generated by the solution, including the node table N(K) and the adjacency matrix A(K), a supplementary resource space K that meets the specifications of the job skills and task process knowledge units is formed; S43-2: Create an S space representation, and include the node table N(K) including instantiated and non-instantiated knowledge entities or nodes consisting of labels and attributes. The [0,1] value in the adjacency matrix A(K) indicates whether there is a connection between the starting entity of the corresponding row and the ending entity of the corresponding column. 0 indicates no connection and 1 indicates a connection.

5. The method for constructing a concept-decision-knowledge job skill map according to claim 1, characterized in that: The S5 specifically includes: S51: Perform demand identification; match job skills and task processes with nodes in the knowledge graph, convert them into feature adjacency matrices with specific structures and connotations according to the CDK space representation, and form an initial demand state space based on deep reinforcement learning; S51-1: Process requirement representation, using the knowledge units of job skills and task processes, decision makers identify through word segmentation and named entity natural language processing technology, extract relevant process concepts from the design requirement text, and form a requirement list consisting of labels and attributes; S51-2: Solution representation, based on the knowledge representation definition of job skills and task processes CDK, the instantiation representation of the initial process requirements is integrated into the process requirements adjacency matrix A0(C); at the same time, the process of generating job skills and task process solutions is abstracted to add instantiated nodes and relationships to the adjacency matrix. Therefore, the features in A0(C) are allocated to the D and K spaces; S51-3: Assign the concepts in the C space to the D and K spaces, and select the minimum subgraph that covers the concepts of the proposed process requirements according to the knowledge unit representation; assign the concepts in the minimum subgraph to the D space to form an adjacency matrix A0(D) representing job skills and task process information; then, assign the concepts that are not in the minimum subgraph to the K space to generate an adjacency matrix A0(K) representing supplementary job skills and task process resources; S52: Carry out model construction and design a deep deterministic policy gradient reinforcement learning algorithm. By jointly training the actor network for learning the policy and the critic network for learning the state value, a deep neural network is used as an approximation of the policy network and the action value function. Stochastic gradient descent is used to train the parameters of the policy network and value network model. A dual neural network is introduced to achieve a more stable learning process. S52-1: Construct the state space, represent the structure and content state of the solution as the state matrix S, and use the adjacency matrix and node table to represent the structure and content of the solution respectively; S52-2: Construct action space, use stored cases or rule knowledge to update the structure and content of job skills and task processes, and add<head,relation,tail> The process of converting the knowledge unit of the triple structure into the state space is transformed into the instance selection of the head entity or the tail entity in the node table, which is used to create a relationship in the non-diagonal position of the adjacency matrix; each action is represented as an action matrix A, forming a matrix with a dimension of A_dim=2*2([Vector h ,Index h ],[Vector t ,Index t ])matrix; S53: Design a reward mechanism that includes regional weights, unit rewards, and outcome evaluation; S53-1: Create regional weights, obtain the connotation of the job skill process by merging the D and K feature spaces, and divide the state space into three regions: D, DK, and K; S53-2: Create unit rewards, divide the reward direction according to the newly added process knowledge after the state update, and divide the action space A t The corresponding triple<head,relation,tail> Find the corresponding case or rule knowledge unit, represented by A t ∈{Case_Units,Rule_Units}); S53-3: Create result evaluation and calculate the state space S for each step t The adjacency matrix A(S t ) and the cosine similarity SS between the adjacency matrix A(PS_KGM) corresponding to the knowledge unit t .

6. The method for constructing a concept-decision-knowledge job skill map according to claim 5, characterized in that: In S52-1, the initial state space S0 of the input model is the initial solution space A(D&K) obtained by demand identification; based on the initial state space, the solution generation process involves updating the state space; the instantiation of each node vector in the state matrix node table represents the addition of one job skill entity to the solution.

Images