Proprietary education knowledge graph skill demand extraction method based on template matching and LLM

Abstract

The invention discloses a template matching and LLM-based production and education knowledge graph skill demand extraction method. The method comprises the steps of S1, obtaining an original text data set related to skill demands; s2, performing preliminary matching on the original text data set by using a predefined template set; s3, identifying and extracting skill demand information which is not captured in template matching; s4, constructing a cross-domain skill demand map; s5, classifying the skill demand information extracted through linkage identification, dividing the skill demand information into different technical fields, industry applications and education demand categories according to a predetermined classification standard, and structuring the skill demand information into a cross-domain skill demand map; and S6, dynamically updating parameters of the template set and the BERT model, and continuously optimizing and expanding a skill demand extraction method according to a linkage recognition result of new text data and cross-domain skill demands. According to the invention, powerful support is provided for multi-field and multi-dimensional skill demand management.

Description

Technical Field

[0001] The present invention relates to the technical field of industry-education knowledge graphs, and in particular to a method for extracting skill requirements from industry-education knowledge graphs based on template matching and LLM. Background Art

[0002] In the field of industry-education integration, accurate identification and structured management of skill requirements are crucial to the effective matching between industrial development and education supply. Existing methods for identifying skill requirements mainly rely on manual rules, simple text classification or keyword matching techniques. The above methods usually show obvious limitations when processing complex and unstructured text data, especially in the case of diverse text expressions and implicit semantics. Traditional technologies find it difficult to accurately identify relevant skill requirements, which often leads to information omissions or misjudgments. In this case, the identification and analysis of skill requirements in industry-education integration becomes more difficult, and it is difficult to effectively support the construction of a skill requirement database, which in turn affects the rational allocation and matching of industrial and educational resources.

[0003] In the existing technology, some studies have attempted to introduce natural language processing technology to improve the accuracy of skill requirement identification. However, traditional natural language processing technology still mainly relies on shallow text feature extraction and rule-based analysis models, which makes it difficult to capture the deep semantics and complex contextual associations in the text. As a result, when faced with large-scale, multi-field skill requirement data, traditional methods cannot effectively cope with the complexity and diversity of the data. In addition, existing skill requirement identification systems usually lack the ability to structured manage and dynamically update skill requirement information, making them unable to cope with rapidly changing industry needs and education supply.

[0004] In terms of structured management of skill demand information, existing knowledge management systems often rely on manually constructed knowledge bases and static classification standards. Although these systems can provide certain reference value in certain specific fields, with industrial development and technological advancement, skill requirements are showing a trend of rapid change and cross-domain intersection. Existing systems are difficult to reflect and adapt to such changes in a timely manner. In addition, the updating and maintenance of the knowledge base mainly rely on manual operations and lack an automated updating mechanism, which makes the knowledge base easily outdated and unable to reflect the latest skill demand information in a timely manner, thereby reducing the practicality and foresight of the system.

[0005] In summary, the existing technologies have the following major defects in the identification, structured management and dynamic updating of skill requirements: First, the existing skill requirement identification methods have limited processing capabilities for complex texts and implicit semantics, which can easily lead to information omissions and misjudgments; second, the existing knowledge management systems lack the ability to dynamically update skill requirement information and manage cross-domain associations, making it difficult to adapt to rapidly changing industry needs and educational supply; finally, the existing skill requirement information structuring methods rely too much on manual rules and lack automated and intelligent processing methods, resulting in insufficient adaptability and practicality of the system. Summary of the invention

[0006] One purpose of the present invention is to propose a method for extracting skill requirements from industry-education knowledge graphs based on template matching and LLM. The present invention provides strong support for multi-field and multi-dimensional skill requirement management.

[0007] According to an embodiment of the present invention, a method for extracting skill requirements from an industry-education knowledge graph based on template matching and LLM includes the following steps:

[0008] S1. Obtain original text datasets related to skill requirements. The original text datasets come from industry reports, educational resource databases, recruitment information, and educational course outlines.

[0009] S2. Perform preliminary matching on the original text dataset using a predefined template set, which is designed based on common features of industries and education fields and can identify specific skill demand patterns in the original text dataset;

[0010] S3. Input the text dataset after preliminary matching into the pre-trained BERT model. Through the deep semantic understanding of the text dataset by the BERT model, identify and extract the skill requirement information that is not captured by template matching.

[0011] S4. When the system identifies the skill demand information in a specific field, it automatically triggers the linkage identification of skill demand in other fields related to the corresponding field. The linkage identification includes analyzing the skill demand in the related field and extracting the relevant skill demand information to construct a cross-field skill demand map.

[0012] S5. Classify the skill demand information extracted through linkage identification, divide the skill demand information into different technical fields, industry applications and educational demand categories according to the predetermined classification standards, and structure the skill demand information into a cross-domain skill demand map, which shows the relationship between skill demands in different fields and the matching degree between industry demands and educational resources;

[0013] S6. Dynamically update the template set and the parameters of the BERT model, and continuously optimize and expand the skill requirement extraction method based on the linkage recognition results of new text data and cross-domain skill requirements.

[0014] Optionally, S1 includes the following sub-steps:

[0015] S11. Collect original text datasets related to skill requirements. The original text datasets include data sets from multiple information sources, denoted as D IR , D EDB , D RI and D ECD :

[0016] D IR Represents a dataset from industry reports, which are published by industry associations, market research companies or government agencies and cover the development trends, technology needs and talent demand information of a specific industry;

[0017] D EDB represents a dataset from an educational resource database, which includes course resources and syllabi from colleges and universities, vocational training institutions, and online education platforms, covering course objectives, teaching content, and skill requirements;

[0018] D RI Represents a dataset from job postings, which are posted on corporate websites, job search platforms, or social media, including the company's skill requirements and job responsibilities for specific positions;

[0019] D ECD Represents a dataset from educational course syllabi, which are published by educational institutions and list the course's teaching objectives, content outlines, learning outcomes, and assessment criteria;

[0020] S12. Preprocess the collected original text dataset, including removing irrelevant characters, punctuation marks and redundant information, and standardizing the text to generate a standardized original text dataset D STD :

[0021] D STD =f pre (D IR ,D EDB ,D RI ,D ECD );

[0022] Among them, f pre For the preprocessing function, the input raw text data set is cleaned, normalized and formatted;

[0023] S13, the preprocessed original text dataset D STDAnnotate text according to its source and generate a standardized original text dataset with source identification

[0024]

[0025] Among them, d i is a single text record in the dataset, l i As the corresponding source labels, IR stands for industry report, EDB stands for education resource database, RI stands for recruitment information, and ECD stands for education course outline.

[0026] Optionally, S2 includes the following sub-steps:

[0027] S21. Design and define a template set T. Each template t in the template set i Based on the common characteristics of the industry and education fields, the template t i Used to identify specific skill demand patterns:

[0028]

[0029] Among them, t i represents the i-th template, k j represents the jth keyword in the set of keywords related to skill requirements, s k represents the kth phrase structure in the phrase structure set, c l represents the lth rule in the context semantic rule set, α j , β k and γ l They are the corresponding weight coefficients, which are used to indicate the influence of different elements on the template matching effect, φ(t i ) is a regularization function used to ensure the overall balance and generalization ability of the template set;

[0030] S22, the preprocessed standardized original text dataset with source identification Match with the template set T to generate a preliminary matching result set M:

[0031]

[0032] Among them, m i represents the i-th matching result in the preliminary matching result set, d j is the jth text record in the standardized original text dataset with source identification, t k is the kth template in the template set, δ jk To match the indicator function, indicate the text record d j Whether to match template t k ,ψ(dj ,t k ) is a matching score function used to evaluate the matching degree between the text record and the template, η jk is the matching weight, which is used to adjust the importance of different templates in the matching process. j | and |t k |The length of the text record and template, respectively;

[0033] S23, filtering and optimizing the preliminary matching result set M, removing duplicate or invalid matches, and generating an optimized matching result set M opt :

[0034]

[0035] in, represents the optimized i-th matching result, λ i is the confidence of the preliminary matching result, θ(m i ) is a deduplication function used to remove duplicate matching results, ω(m i ) is the invalid matching filter function, and |M| is the size of the preliminary matching result set.

[0036] Optionally, S3 includes the following sub-steps:

[0037] S31, the text data set M after preliminary matching opt The data is input into the pre-trained BERT model. The BERT model extracts features from the input text data based on the pre-trained deep semantic understanding ability and generates a semantic feature vector V. sem ;

[0038] S32, through the semantic feature vector V sem Matching result M with the original template opt Perform interactive calculations to identify implicit skill requirement information that is not captured by template matching and generate a supplementary matching result set M supp :

[0039]

[0040] in, represents the i-th supplementary matching result in the supplementary matching result set, |V sem | represents the semantic feature vector V sem The dimension represents the complexity of the semantic information generated by the BERT model. Tanh is the hyperbolic tangent function, which is used to introduce nonlinear characteristics to help the model better capture complex semantic relationships. is the initial matching result set M opt The feature dimension of the i-th matching result in represents the information related to template matching, β jkis the interaction weight between the j-th dimension semantic feature and the k-th dimension preliminary matching result, indicating the strength of the association between the two in identifying implicit skill requirement information. is the regularization factor to ensure the stability of interactive calculations and the balance of results. Represents semantic feature vector With the preliminary matching results The cosine similarity between them is used to quantify the similarity between them and capture the potential skill requirements, ψ jk The matching adjustment coefficient is used to dynamically adjust the matching weights between different features to optimize the accuracy of the supplementary matching results;

[0041] S33, supplement the matching result set M supp and the preliminary matching result set M opt Merge to generate the final matching result set M final :

[0042] M final =M opt ∪M supp ;

[0043] Among them, M final The final skill requirement matching result set includes preliminary matching results and supplementary matching results identified by the BERT model.

[0044] Optionally, S4 includes the following sub-steps:

[0045] S41. When the system identifies the skill demand information in a specific field After that, the linkage recognition module is automatically triggered. The linkage recognition module is based on the predefined domain association matrix A field Identify other areas that are relevant to a particular area;

[0046] S42, based on the domain association matrix A field The results are determined by the identified field x Related related domain set R field :

[0047] R field ={field j |a xj >θandfield j ≠field x};

[0048] Among them, R field Representation and field x A set of related fields, θ is the correlation threshold, which is used to filter out fields with correlation exceeding the threshold;

[0049] S43, for the associated domain set R field Each field in j Conduct linkage identification and extract relevant skill demand information Generate linkage recognition result set M link :

[0050]

[0051] in, For the associated field field j The skill requirement information extracted through linkage identification in function It is used to combine the final matching results and related domain features for linkage analysis, and extract cross-domain skill requirements related to the original domain skill requirement information;

[0052] S44, linking the recognition result set M link The final matching result set M with the original recognition domain final Integrate and build a cross-domain skill demand mapG cross .

[0053] Optionally, the S44 includes the following sub-steps:

[0054] S441, linking the recognition result set M link The final matching result set M with the original recognition domain final Perform data fusion to generate a preliminary cross-domain skill requirement set M cross ;

[0055] S442. Preliminary cross-domain skills requirements set M cross Conduct relationship analysis and establish the correlation matrix R between skill requirements cross ;

[0056] S443, based on the correlation matrix R cross , the cross-domain skills set M cross Construct a cross-domain skill demand map and generate a cross-domain skill demand map G cross :

[0057] G cross =(V cross ,E cross );

[0058] Among them, G cross represents the final cross-domain skill demand graph, V cross is the node set of the cross-domain skill demand graph, representing the information of each skill demand. cross is the edge set of the cross-domain skill demand graph, indicating the relationship between skill requirements:

[0059]

[0060] Among them, e ij Represents node m i and node m j The edge between is an edge in the cross-domain skill demand graph, reflecting the association between two skill requirements. x Represents the domain node set V cross The weight of the xth skill requirement node in reflects its importance in the cross-domain skill requirement graph, λ x is a regularization parameter used to control the node m i and node m j The effect of the distance between them on the edge weight, the higher the λ x The value makes the edge weights between nodes with a longer distance lower, thus ensuring that the construction of the edges in the cross-domain skill demand graph follows the distance constraint and ensures the accuracy of the cross-domain skill demand graph. i ,m j ) represents node m i and node m j The semantic distance between them is based on the semantic feature vector V of the skill requirement. sem The smaller the distance is, the stronger the semantic association between the two skill requirements is, and the stronger the edge between the two in the cross-domain skill requirement graph is. x Represents the node m i and node m j The associated domain-specific weights are used to adjust the weights of the skill demand edges in a specific domain, thereby ensuring that the construction of the cross-domain skill demand map fully considers the domain specificity. Represents node m i and node m j The semantic similarity between them is calculated by calculating the semantic feature vector V sem The similarity directly affects the weight of the edge in the cross-domain skill demand graph, ensuring that similar skill requirements are closely connected in the cross-domain skill demand graph. xy To adjust the parameters, it means node m i and node m j The influence of the correlation between other nodes on the current edge construction ensures that the edge in the cross-domain skill demand graph not only considers the direct relationship between the two nodes, but also considers their relative position and relationship in the entire cross-domain skill demand graph. i ,m y ) and Rel(m j ,m y ) represents node m i and node m jand other nodes m in the cross-domain skill demand graph y The relationship degree is used to adjust e ij The calculation of φ ensures that the construction of edges takes into account both local associations and global structures. ij The regularization factor is used to ensure that when constructing a cross-domain skill demand graph, the edges in the cross-domain skill demand graph can balance the complex relationships between different nodes and avoid the cross-domain skill demand graph being too dense or sparse.

[0061] Optionally, S5 includes the following sub-steps:

[0062] S51, skill requirement information M extracted through linkage identification link Carry out preliminary classification and sort the skill demand information according to the predetermined classification standards Divided into different technology areas C tech 、Industry Application Collection C ind and the educational demand set C edu ;

[0063] S52, combine technical fields into C tech 、Industry Application Collection C ind and the educational demand set C edu Perform semantic analysis and contextual analysis on the skill requirement information in each classification set to generate a result set of technical fields for refined classification Industry application result set and educational requirements result set

[0064] S53: Structuring the classified skill demand information into the final cross-domain skill demand map based on the cross-domain skill demand map Used to demonstrate the correlation between skill requirements in different fields and the match between industry needs and educational resources:

[0065]

[0066] in, Represents the final cross-domain skill demand map, It is the node set of the final cross-domain skill demand map, representing each refined skill demand information. It is the edge set of the final cross-domain skill requirement graph, representing the association relationship between different skill requirement information.

[0067] The beneficial effects of the present invention are:

[0068] (1) The present invention designs and applies a set of templates based on common features of industries and education fields, and effectively identifies the skill requirement patterns in the text in the initial matching stage. Subsequently, the pre-trained BERT model is used to perform deep semantic analysis on the preliminary matching results, which makes up for the shortcomings of template matching in processing complex and implicit semantics, and can identify skill requirement information that the template fails to capture. Through a multi-level recognition mechanism, the present invention significantly improves the accuracy and comprehensiveness of the extraction of skill requirement information, ensuring the comprehensive capture of important information in complex texts.

[0069] (2) After identifying the skill requirement information in a certain field, the present invention automatically triggers the identification of skill requirements in other fields related to the field through a linkage identification module, constructs a cross-field skill requirement map, and effectively captures and displays the skill requirement associations between different fields through a field association matrix and a cross-field analysis algorithm, helping to establish a more accurate match between industry and educational resources. This not only improves the system's comprehensive analysis capabilities for skill requirement information, but also provides strong support for multi-field and multi-dimensional skill requirement management.

[0070] (3) The present invention significantly improves the efficiency of organizing and managing skill requirement information by classifying and refining the identified skill requirement information and further structuring it into a cross-domain skill requirement map. Through predefined classification standards and semantic analysis algorithms, the system can accurately classify skill requirement information and generate a clear knowledge map to show the correlation between various skill requirements. In addition, the introduction of a dynamic update mechanism can automatically optimize and update the template set and BERT model parameters according to the newly identified skill requirement information, ensuring that the system can always reflect the latest industry needs and education trends.

[0071] (4) When constructing a cross-domain skill requirement map, the present invention introduces multi-level association analysis and map optimization algorithms, so that the final generated skill requirement map not only accurately displays the association between skill requirements, but also has a good visualization effect. By optimizing parameters such as the weights, semantic distances and similarities of map nodes and edges, the skill requirement information is intuitively displayed, which facilitates users to understand and use the skill requirement map for analysis and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0073] Figure 1 This is a flow chart of a method for extracting skill requirements from an industry-education knowledge graph based on template matching and LLM proposed by the present invention;

[0074] Figure 2This is a flowchart of the linkage identification and cross-domain skill requirement map construction described in step S4 of the method for extracting skill requirements from the industry-education knowledge graph based on template matching and LLM proposed by the present invention. DETAILED DESCRIPTION

[0075] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0076] refer to Figure 1-2 , a method for extracting skill requirements from industry-education knowledge graph based on template matching and LLM, comprising the following steps:

[0077] S1. Obtain original text datasets related to skill requirements. The original text datasets come from industry reports, educational resource databases, recruitment information, and educational course outlines.

[0078] S2. Perform preliminary matching on the original text dataset using a predefined template set, which is designed based on common features of industries and education fields and can identify specific skill demand patterns in the original text dataset;

[0079] S3. Input the text dataset after preliminary matching into the pre-trained BERT model. Through the deep semantic understanding of the text dataset by the BERT model, identify and extract the skill requirement information that is not captured by template matching.

[0080] S4. When the system identifies the skill demand information in a specific field, it automatically triggers the linkage identification of skill demand in other fields related to the corresponding field. The linkage identification includes analyzing the skill demand in the related field and extracting the relevant skill demand information to construct a cross-field skill demand map.

[0081] S5. Classify the skill demand information extracted through linkage identification, divide the skill demand information into different technical fields, industry applications and educational demand categories according to the predetermined classification standards, and structure the skill demand information into a cross-domain skill demand map, which shows the relationship between skill demands in different fields and the matching degree between industry demands and educational resources;

[0082] S6. Dynamically update the template set and the parameters of the BERT model, and continuously optimize and expand the skill requirement extraction method based on the linkage recognition results of new text data and cross-domain skill requirements.

[0083] In this implementation, S1 includes the following sub-steps:

[0084] S11. Collect original text datasets related to skill requirements. The original text datasets include data sets from multiple information sources, denoted as D IR , D EDB , D RI and D ECD :

[0085] D IR Represents a dataset from industry reports, which are published by industry associations, market research companies or government agencies and cover the development trends, technology needs and talent demand information of a specific industry;

[0086] D EDB represents a dataset from an educational resource database, which includes course resources and syllabi from colleges and universities, vocational training institutions, and online education platforms, covering course objectives, teaching content, and skill requirements;

[0087] D RI Represents a dataset from job postings, which are posted on corporate websites, job search platforms, or social media, including the company's skill requirements and job responsibilities for specific positions;

[0088] D ECD Represents a dataset from educational course syllabi, which are published by educational institutions and list the course's teaching objectives, content outlines, learning outcomes, and assessment criteria;

[0089] S12. Preprocess the collected original text dataset, including removing irrelevant characters, punctuation marks and redundant information, and standardizing the text to generate a standardized original text dataset D STD :

[0090] D STD =f pre (D IR ,D EDB ,D RI ,D ECD );

[0091] Among them, f pre For the preprocessing function, the input raw text data set is cleaned, normalized and formatted;

[0092] S13, the preprocessed original text dataset D STD Annotate text according to its source and generate a standardized original text dataset with source identification

[0093]

[0094] Among them, d i is a single text record in the dataset, li As the corresponding source labels, IR stands for industry report, EDB stands for education resource database, RI stands for recruitment information, and ECD stands for education course outline.

[0095] In this implementation, S2 includes the following sub-steps:

[0096] S21. Design and define a template set T. Each template t in the template set i Based on the common characteristics of the industry and education fields, the template t i Used to identify specific skill demand patterns:

[0097]

[0098] Among them, t i represents the i-th template, k j represents the jth keyword in the set of keywords related to skill requirements, s k represents the kth phrase structure in the phrase structure set, c l represents the lth rule in the context semantic rule set, α j , β k and γ l They are the corresponding weight coefficients, which are used to indicate the influence of different elements on the template matching effect, φ(t i ) is a regularization function used to ensure the overall balance and generalization ability of the template set;

[0099] S22, the preprocessed standardized original text dataset with source identification Match with the template set T to generate a preliminary matching result set M:

[0100]

[0101] Among them, m i represents the i-th matching result in the preliminary matching result set, d j is the jth text record in the standardized original text dataset with source identification, t k is the kth template in the template set, δ jk To match the indicator function, indicate the text record d j Whether to match template t k ,ψ(d j ,t k ) is a matching score function used to evaluate the matching degree between the text record and the template, η jk is the matching weight, which is used to adjust the importance of different templates in the matching process. j | and |t k |The length of the text record and template, respectively;

[0102] S23, filtering and optimizing the preliminary matching result set M, removing duplicate or invalid matches, and generating an optimized matching result set M opt :

[0103]

[0104] in, represents the optimized i-th matching result, λ i is the confidence of the preliminary matching result, θ(m i ) is a deduplication function used to remove duplicate matching results, ω(m i ) is the invalid matching filter function, and |M| is the size of the preliminary matching result set.

[0105] In this implementation, S3 includes the following sub-steps:

[0106] S31, the text data set M after preliminary matching opt The data is input into the pre-trained BERT model. The BERT model extracts features from the input text data based on the pre-trained deep semantic understanding ability and generates a semantic feature vector V. sem ;

[0107] S32, through the semantic feature vector V sem Matching result M with the original template opt Perform interactive calculations to identify implicit skill requirement information that is not captured by template matching and generate a supplementary matching result set M supp :

[0108]

[0109] in, represents the i-th supplementary matching result in the supplementary matching result set, |V sem | represents the semantic feature vector V sem The dimension represents the complexity of the semantic information generated by the BERT model. Tanh is the hyperbolic tangent function, which is used to introduce nonlinear characteristics to help the model better capture complex semantic relationships. is the initial matching result set M opt The feature dimension of the i-th matching result in represents the information related to template matching, β jk is the interaction weight between the j-th dimension semantic feature and the k-th dimension preliminary matching result, indicating the strength of the association between the two in identifying implicit skill requirement information. is the regularization factor to ensure the stability of interactive calculations and the balance of results. Represents semantic feature vector With the preliminary matching results The cosine similarity between them is used to quantify the similarity between them and capture the potential skill requirements, ψ jk The matching adjustment coefficient is used to dynamically adjust the matching weights between different features to optimize the accuracy of the supplementary matching results;

[0110] S33, supplement the matching result set M supp and the preliminary matching result set M opt Merge to generate the final matching result set M final :

[0111] M final =M opt ∪M supp ;

[0112] Among them, M final The final skill requirement matching result set includes preliminary matching results and supplementary matching results identified by the BERT model.

[0113] In this implementation, S4 includes the following sub-steps:

[0114] S41. When the system identifies the skill demand information in a specific field After that, the linkage recognition module is automatically triggered. The linkage recognition module is based on the predefined domain association matrix A field Identify other areas that are relevant to a particular area;

[0115] S42, based on the domain association matrix A field The results are determined by the identified field x Related related domain set R field :

[0116] R field ={field j |a xj >θandfield j ≠field x};

[0117] Among them, R field Representation and field x A set of related fields, θ is the correlation threshold, which is used to filter out fields with correlation exceeding the threshold;

[0118] S43, for the associated domain set R field Each field in j Conduct linkage identification and extract relevant skill demand information Generate linkage recognition result set M link :

[0119]

[0120] in, For the associated field field j The skill requirement information extracted through linkage identification in function It is used to combine the final matching results and related domain features for linkage analysis, and extract cross-domain skill requirements related to the original domain skill requirement information;

[0121] S44, linking the recognition result set M link The final matching result set M with the original recognition domain final Integrate and build a cross-domain skill demand mapG cross .

[0122] In this implementation, S44 includes the following sub-steps:

[0123] S441, linking the recognition result set M link The final matching result set M with the original recognition domain final Perform data fusion to generate a preliminary cross-domain skill requirement set M cross ;

[0124] S442. Preliminary cross-domain skills requirements set M cross Conduct relationship analysis and establish the correlation matrix R between skill requirements cross ;

[0125] S443, based on the correlation matrix R cross , the cross-domain skills set M cross Construct a cross-domain skill demand map and generate a cross-domain skill demand map G cross :

[0126] G cross =(V cross ,E cross );

[0127] Among them, G cross represents the final cross-domain skill demand graph, V cross is the node set of the cross-domain skill demand graph, representing the information of each skill demand. cross is the edge set of the cross-domain skill demand graph, indicating the relationship between skill requirements:

[0128]

[0129] Among them, e ij Represents node m i and node m j The edge between is an edge in the cross-domain skill demand graph, reflecting the association between two skill requirements.x Represents the domain node set V cross The weight of the xth skill requirement node in reflects its importance in the cross-domain skill requirement graph, λ x is a regularization parameter used to control the node m i and node m j The effect of the distance between them on the edge weight, the higher the λ x The value makes the edge weights between nodes with a longer distance lower, thus ensuring that the construction of the edges in the cross-domain skill demand graph follows the distance constraint and ensures the accuracy of the cross-domain skill demand graph. i ,m j ) represents node m i and node m j The semantic distance between them is based on the semantic feature vector V of the skill requirement. sem The smaller the distance is, the stronger the semantic association between the two skill requirements is, and the stronger the edge between the two in the cross-domain skill requirement graph is. x Represents the node m i and node m j The associated domain-specific weights are used to adjust the weights of the skill demand edges in a specific domain, thereby ensuring that the construction of the cross-domain skill demand map fully considers the domain specificity. Represents node m i and node m j The semantic similarity between them is calculated by calculating the semantic feature vector V sem The similarity directly affects the weight of the edge in the cross-domain skill demand graph, ensuring that similar skill requirements are closely connected in the cross-domain skill demand graph. xy To adjust the parameters, it means node m i and node m j The influence of the correlation between other nodes on the current edge construction ensures that the edge in the cross-domain skill demand graph not only considers the direct relationship between the two nodes, but also considers their relative position and relationship in the entire cross-domain skill demand graph. i ,m y ) and Rel(m j ,m y ) represents node m i and node m j and other nodes m in the cross-domain skill demand graph y The relationship degree is used to adjust e ij The calculation of φ ensures that the construction of edges takes into account both local associations and global structures. ijThe regularization factor is used to ensure that when constructing a cross-domain skill demand graph, the edges in the cross-domain skill demand graph can balance the complex relationships between different nodes and avoid the cross-domain skill demand graph being too dense or sparse.

[0130] In this implementation, S5 includes the following sub-steps:

[0131] S51, skill requirement information M extracted through linkage identification link Carry out preliminary classification and sort the skill demand information according to the predetermined classification standards Divided into different technology areas C tech 、Industry Application Collection C ind and the educational demand set C edu ;

[0132] S52, combine technical fields into C tech 、Industry Application Collection C ind and the educational demand set C edu Perform semantic analysis and contextual analysis on the skill requirement information in each classification set to generate a result set of technical fields for refined classification Industry application result set and educational requirements result set

[0133] S53: Structuring the classified skill demand information into the final cross-domain skill demand map based on the cross-domain skill demand map Used to demonstrate the correlation between skill requirements in different fields and the match between industry needs and educational resources:

[0134]

[0135] in, Represents the final cross-domain skill demand map, It is the node set of the final cross-domain skill demand map, representing each refined skill demand information. It is the edge set of the final cross-domain skill requirement graph, representing the association relationship between different skill requirement information.

[0136] Embodiment 1:

[0137] In a smart manufacturing industrial park in a certain province, with the rapid development of artificial intelligence technology, enterprises have an increasing demand for senior engineers with AI development and big data processing capabilities. At the same time, local colleges and vocational training institutions are also stepping up to adjust the course content to better meet the needs of the industry. However, due to the information asymmetry between the industry and the education field, the skill demand information released by the enterprise is often difficult to be accurately identified and fed back to the educational institutions, resulting in a mismatch between educational resources and market demand, affecting the efficiency and quality of talent training. A smart manufacturing company X recently released a series of recruitment needs, covering multiple technical fields such as artificial intelligence, big data processing, and the Internet of Things. However, in the recruitment information, the description of skill requirements is diverse and the implicit semantics are complex, which brings challenges to the accurate identification of skill requirements by educational institutions. In order to better solve this problem, the education alliance of the industrial park decided to introduce the method of the present invention, by constructing an industry-education knowledge graph, accurately identifying the skill requirements of enterprises, and guiding colleges and vocational training institutions to adjust the course settings to improve the accuracy of talent training.

[0138] The job postings published by Company X include the following skill requirement descriptions for the following positions: “We need an AI algorithm engineer who is proficient in TensorFlow and can handle large-scale data sets. He / she must have some experience in IoT applications and be able to design and optimize neural network models. In addition, the position also requires a deep understanding of big data analysis and cloud computing.” The requirements cover skills in multiple fields, including both explicitly mentioned skills (TensorFlow, neural network model design) and implicit related requirements (IoT, big data analysis, cloud computing). Traditional skill requirement identification methods, such as keyword-based matching, may miss implicit skill requirements, resulting in incomplete identification results.

[0139] In this embodiment, the education alliance uses the method of the present invention to extract and analyze skill requirements. First, the system collects all recruitment information released by Company X and preprocesses the text data, including removing redundant information and standardizing the text format. Then, the system uses a predefined template set to perform a preliminary match on the preprocessed text data. The template set is designed based on the common features of the intelligent manufacturing field and can efficiently identify the explicit skill requirements in the text, such as "TensorFlow", "AI algorithm", and "big data processing".

[0140] After the initial matching, the system inputs the text data into the pre-trained BERT model. Through the deep semantic understanding of the BERT model, it further identifies implicit skill requirement information, such as "IoT application experience" and "cloud computing", which are not directly mentioned in the text but are closely related to clear skill requirements. The BERT model successfully extracts implicit requirement information based on the contextual semantic relationship, making up for the shortcomings of the template matching stage.

[0141] When the system identifies X company’s skill needs in the fields of artificial intelligence and big data processing, the linkage identification module is automatically triggered, and linkage identification of skill needs in the fields of the Internet of Things and cloud computing is performed. The system determines the correlation between fields based on the predefined field association matrix, and further extracts other field skills related to the skill needs, such as “edge computing” and “sensor technology”. Skills are often ignored in traditional methods, but are accurately identified and processed in the method of the present invention.

[0142] Through linkage identification, the system not only successfully extracted cross-domain skill requirements, but also structured the demand information into a cross-domain skill demand map. The map clearly shows the relationship between Company X’s skill requirements in the fields of artificial intelligence, big data, and the Internet of Things, as well as the matching degree between the requirements and existing educational resources. The map provides an intuitive perspective, enabling educational institutions to quickly identify course content that needs to be adjusted and formulate corresponding education plans.

[0143] Before the implementation of the present invention, the Education Alliance had tried traditional skill requirement identification methods, such as keyword matching and rule analysis. In order to compare the effects of the method of the present invention and the traditional method, the Education Alliance compared the performance of the two on the same data set. The specific training samples included 500 recruitment information released by Company X, covering multiple technical fields such as intelligent manufacturing, artificial intelligence, and big data. In the experiment, the Education Alliance used the traditional method and the method of the present invention to process the data, and recorded their respective recognition results, accuracy, underreporting rate, and processing time indicators.

[0144] The experimental results show that the traditional method identified a total of 320 skill requirement information, with an average accuracy of 68%, an underreporting rate of 25%, and a total processing time of 2 hours. The traditional method performed well in identifying explicit skill requirements (explicitly mentioned skill vocabulary), but its ability to identify implicit requirements and cross-domain related requirements was obviously insufficient, and the underreporting rate was high. In addition, the processing time of the traditional method was relatively long, mainly because it required manual intervention and repeated analysis.

[0145] In comparison, the method of the present invention identified a total of 470 pieces of skill requirement information, with an average accuracy of 92%, an underreporting rate of only 8%, and a total processing time of 1 hour. In the identification of explicit skill requirements, the method of the present invention has a higher accuracy rate and can effectively identify implicit requirement information that traditional methods fail to capture. In addition, through the linkage recognition module, the method of the present invention can automatically identify and process cross-domain skill requirements, greatly improving the comprehensiveness and accuracy of identification.

[0146] Through data comparison, it can be clearly seen that the method of the present invention is significantly superior to traditional methods in terms of accuracy, comprehensiveness and processing efficiency in identifying skill requirements. In particular, when dealing with complex and multi-domain cross-skill requirements in the field of intelligent manufacturing, the method of the present invention can give full play to its technical advantages and provide a more accurate and efficient solution. The improvement not only significantly improves the work efficiency of the education alliance, but also provides strong data support for local colleges and vocational training institutions, helping them to better adapt to changes in market demand and formulate more accurate talent training programs.

[0147] The present invention designs and applies a set of templates based on common features in industries and education fields to effectively identify skill requirement patterns in texts in the initial matching stage. Subsequently, the pre-trained BERT model is used to perform deep semantic analysis on the initial matching results, which makes up for the shortcomings of template matching in processing complex and implicit semantics, and can identify skill requirement information that the template fails to capture. Through a multi-level recognition mechanism, the present invention significantly improves the accuracy and comprehensiveness of the extraction of skill requirement information, ensuring the comprehensive capture of important information in complex texts.

[0148] After identifying the skill requirement information in a certain field, the present invention automatically triggers the identification of skill requirements in other fields related to the field through a linkage identification module, constructs a cross-field skill requirement map, and effectively captures and displays the skill requirement associations between different fields through a field association matrix and a cross-field analysis algorithm, helping to establish a more accurate match between industry and educational resources. It not only improves the system's comprehensive analysis capabilities for skill requirement information, but also provides strong support for multi-field and multi-dimensional skill requirement management.

[0149] The present invention significantly improves the efficiency of organizing and managing skill requirement information by classifying and refining the identified skill requirement information and further structuring it into a cross-domain skill requirement map. Through predefined classification standards and semantic analysis algorithms, the system can accurately classify skill requirement information and generate a clear knowledge map to show the correlation between various skill requirements. In addition, the introduction of a dynamic update mechanism can automatically optimize and update the template set and BERT model parameters according to the newly identified skill requirement information, ensuring that the system can always reflect the latest industry needs and education trends.

[0150] When constructing a cross-domain skill requirement map, the present invention introduces multi-level association analysis and map optimization algorithms, so that the final generated skill requirement map not only accurately displays the association between skill requirements, but also has a good visualization effect. By optimizing parameters such as the weights, semantic distances and similarities of map nodes and edges, the skill requirement information is intuitively displayed, which is convenient for users to understand and use the skill requirement map for analysis and decision-making.

[0151] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for extracting skill requirements from industry-education knowledge graph based on template matching and LLM, characterized in that: The steps include: S1. Obtain original text datasets related to skill requirements. The original text datasets come from industry reports, educational resource databases, recruitment information, and educational course outlines. S2. Perform preliminary matching on the original text dataset using a predefined template set, where the template set is designed based on common features of industries and education fields to identify specific skill demand patterns in the original text dataset; S3. Input the text dataset after preliminary matching into the pre-trained BERT model. Through the deep semantic understanding of the text dataset by the BERT model, identify and extract the skill requirement information that is not captured by template matching. S4. When the system identifies the skill demand information in a specific field, it automatically triggers the linkage identification of skill demand in other fields related to the corresponding field. The linkage identification includes analyzing the skill demand in the related field, extracting the relevant skill demand information, and building a cross-field skill demand map. S5. Classify the skill demand information extracted through linkage identification, divide the skill demand information into different technical fields, industry applications and educational demand categories according to the predetermined classification standards, and structure the skill demand information into a cross-domain skill demand map, which shows the relationship between skill demands in different fields and the matching degree between industry demands and educational resources; S6. Dynamically update the template set and the parameters of the BERT model, and continuously optimize and expand the skill requirement extraction method based on the linkage recognition results of new text data and cross-domain skill requirements.

2. According to claim 1, a method for extracting skill requirements from industry-education knowledge graph based on template matching and LLM is characterized in that: The S1 comprises the following sub-steps: S11. Collect original text datasets related to skill requirements. The original text datasets include data sets from multiple information sources, denoted as D IR , D EDB , D RI and D ECD : D IR Represents a dataset from industry reports, which are published by industry associations, market research companies or government agencies and cover the development trends, technology needs and talent demand information of a specific industry; D EDB represents a dataset from an educational resource database, which includes course resources and syllabi from colleges and universities, vocational training institutions, and online education platforms, covering course objectives, teaching content, and skill requirements; D RI Represents a dataset from job postings, which are posted on corporate websites, job search platforms, or social media, including the company's skill requirements and job responsibilities for specific positions; D ECD Represents a dataset from educational course syllabi, which are published by educational institutions and list the course's teaching objectives, content outlines, learning outcomes, and assessment criteria; S12. Preprocess the collected original text dataset, including removing irrelevant characters, punctuation marks and redundant information, and standardizing the text to generate a standardized original text dataset D STD : D STD =f pre (D IR ,D EDB ,D RI ,D ECD ); Among them, f pre For the preprocessing function, the input raw text data set is cleaned, normalized and formatted; S13, the preprocessed original text dataset D STD Annotate text according to its source and generate a standardized original text dataset with source identification Among them, d i is a single text record in the dataset, l i As the corresponding source labels, IR stands for industry report, EDB stands for education resource database, RI stands for recruitment information, and ECD stands for education course outline.

3. According to claim 1, a method for extracting skill requirements from industry-education knowledge graph based on template matching and LLM is characterized in that: The S2 comprises the following sub-steps: S21. Design and define a template set T. Each template t in the template set i Based on the common characteristics of the industry and education fields, the template t i Used to identify specific skill demand patterns: Among them, t i represents the i-th template, k j represents the jth keyword in the set of keywords related to skill requirements, s k represents the kth phrase structure in the phrase structure set, c l represents the lth rule in the context semantic rule set, α j , β k and γ l They are the corresponding weight coefficients, which are used to indicate the influence of different elements on the template matching effect, φ(t i ) is the regularization function; S22, the preprocessed standardized original text dataset with source identification Match with the template set T to generate a preliminary matching result set M: Among them, m i represents the i-th matching result in the preliminary matching result set, d j is the jth text record in the standardized original text dataset with source identification, t k is the kth template in the template set, δ jk To match the indicator function, indicate the text record d j Whether to match template t k ,ψ(d j ,t k ) is a matching score function used to evaluate the matching degree between the text record and the template, η jk is the matching weight, |d j | and |t k |The length of the text record and template, respectively; S23, filtering and optimizing the preliminary matching result set M, removing duplicate or invalid matches, and generating an optimized matching result set M opt : in, represents the optimized i-th matching result, λ i is the confidence of the preliminary matching result, θ(m i ) is a deduplication function used to remove duplicate matching results, ω(m i ) is the invalid matching filter function, and |M| is the size of the preliminary matching result set.

4. According to claim 1, a method for extracting skill requirements from industry-education knowledge graph based on template matching and LLM is characterized in that: The S3 comprises the following sub-steps: S31, the text data set M after preliminary matching opt The data is input into the pre-trained BERT model. The BERT model extracts features from the input text data based on the pre-trained deep semantic understanding ability and generates a semantic feature vector V. sem ; S32, through the semantic feature vector V sem Matching result M with the original template opt Perform interactive calculations to identify implicit skill requirement information that is not captured by template matching and generate a supplementary matching result set M supp : in, represents the i-th supplementary matching result in the supplementary matching result set, |V sem | represents the semantic feature vector V sem dimension, tanh is the hyperbolic tangent function, which is used to introduce nonlinear characteristics. is the initial matching result set M opt The feature dimension of the i-th matching result in represents the information related to template matching, β jk is the interaction weight between the j-th dimension semantic feature and the k-th dimension preliminary matching result, indicating the strength of the association between the two in identifying implicit skill requirement information. is the regularization factor, Represents semantic feature vector With the preliminary matching results The cosine similarity between them is used to quantify the similarity between them and capture the potential skill requirements, ψ jk It is the matching adjustment coefficient, which is used to dynamically adjust the matching weights between different features; S33, supplement the matching result set M supp and the preliminary matching result set M opt Merge to generate the final matching result set M final : M final =M opt ∪M supp ; Among them, M final The final skill requirement matching result set includes preliminary matching results and supplementary matching results identified by the BERT model.

5. According to claim 1, a method for extracting skill requirements from industry-education knowledge graph based on template matching and LLM is characterized in that: The S4 comprises the following sub-steps: S41. When the system identifies the skill demand information in a specific field After that, the linkage recognition module is automatically triggered. The linkage recognition module is based on the predefined domain association matrix A field Identify other areas that are relevant to a particular area; S42, based on the domain association matrix A field The results are determined by the identified field x Related related domain set R field : R field ={field j |a xj >θand field j ≠field x }; Among them, R field Representation and field x A set of related fields, θ is the correlation threshold, which is used to filter out fields with correlation exceeding the threshold; S43, for the associated domain set R field Each field in j Conduct linkage identification and extract relevant skill demand information Generate linkage recognition result set M link : in, For the associated field field j The skill requirement information extracted through linkage identification in function It is used to combine the final matching results and related domain features for linkage analysis, and extract cross-domain skill requirements related to the original domain skill requirement information; S44, linking the recognition result set M link The final matching result set M with the original recognition domain final Integrate and build a cross-domain skill demand mapG cross .

6. According to claim 5, a method for extracting skill requirements from industry-education knowledge graph based on template matching and LLM is characterized in that: The S44 comprises the following sub-steps: S441, linking the recognition result set M link The final matching result set M with the original recognition domain final Perform data fusion to generate a preliminary cross-domain skill requirement set M cross ; S442. Preliminary cross-domain skills requirements set M cross Conduct relationship analysis and establish the correlation matrix R between skill requirements cross ; S443, based on the correlation matrix R cross , the cross-domain skills set M cross Construct a cross-domain skill demand map and generate a cross-domain skill demand map G cross : G cross =(V cross ,E cross ); Among them, G cross represents the final cross-domain skill demand graph, V cross is the node set of the cross-domain skill demand graph, representing the information of each skill demand. cross is the edge set of the cross-domain skill demand graph, indicating the relationship between skill requirements: Among them, e ij Represents node m i and node m j The edge between is an edge in the cross-domain skill demand graph, reflecting the association between two skill requirements. x Represents the domain node set V cross The weight of the x-th skill requirement node in , λ x is a regularization parameter used to control the node m i and node m j The influence of the distance between them on the edge weight, d(m i ,m j ) represents node m i and node m j The semantic distance between them is based on the semantic feature vector V of the skill requirement. sem Calculated, ω x Represents the node m i and node m j The associated domain-specific weights are used to adjust the weights of the skill demand edges within a specific domain. Represents node m i and node m j The semantic similarity between xy is the adjustment parameter, indicating that node m i and node m j The influence of the association degree with other nodes on the current edge construction, Rel(m i ,m y ) and Rel(m j ,m y ) represents node m i and node m j and other nodes m in the cross-domain skill demand graph y The relationship degree is used to adjust e ij Calculation of φ ij Regularization factor.

7. The method for extracting skill requirements from industry-education knowledge graph based on template matching and LLM according to claim 6 is characterized in that: The S5 comprises the following sub-steps: S51, skill requirement information M extracted through linkage identification link Carry out preliminary classification and sort the skill demand information according to the predetermined classification standards Divided into different technology areas C tech 、Industry Application Collection C ind and the educational demand set C edu ; S52, combine technical fields into C tech 、Industry Application Collection C ind and the educational demand set C edu Perform semantic analysis and contextual analysis on the skill requirement information in each classification set to generate a result set of technical fields for refined classification Industry application result set and educational requirements result set S53: Structuring the classified skill demand information into the final cross-domain skill demand map based on the cross-domain skill demand map Used to demonstrate the correlation between skill requirements in different fields and the match between industry needs and educational resources: in, Represents the final cross-domain skill demand map, It is the node set of the final cross-domain skill demand map, representing each refined skill demand information. It is the edge set of the final cross-domain skill requirement graph, representing the association relationship between different skill requirement information.

Images