Artificial intelligence-based employee skill map construction system

Abstract

The invention discloses an employee skill atlas construction system based on artificial intelligence, and relates to the technical field of atlas construction.The employee skill atlas construction system comprises the steps that firstly, multi-source skill data are obtained, a data processing algorithm based on deep learning is adopted to process the multi-source skill data to mine skill phrases in the multi-source skill data; and further coding the skill phrases into a high-dimensional semantic space, capturing a cross-skill description semantic cooperation relationship in combination with a semantic interaction mechanism, and realizing semantic similarity discrimination of the skill phrases based on the semantic cooperation relationship between the skill phrases. According to the method, a standardized skill system with strong semantic generalization ability is constructed, so that the defect of a traditional method in the aspect of semantic ambiguity elimination of skill data is overcome, the accuracy of employee skill portrait construction is improved, and accurate decision support is provided for an enterprise talent strategy.

Description

Technical Field

[0001] The present application relates to the technical field of graph construction, and more specifically, to an employee skill graph construction system based on artificial intelligence. Background Art

[0002] In existing employee skill management systems, the diversity and semantic ambiguity of skill descriptions have become a core bottleneck hindering the construction of skill graphs. Traditional methods rely on manually pre-defined standardized skill dictionaries, forcing skill phrases to be standardized through keyword matching or regular expressions. However, such rule engines struggle to cope with scenarios where semantically equivalent but heterogeneous expressions exist, and skill entities can easily split due to surface vocabulary differences.

[0003] While existing technologies attempt to incorporate word vector similarity calculations (such as Word2Vec and FastText) for semantic matching, their single-vector global representations cannot distinguish between conflicting ambiguities in context-dependent skill descriptions and lack fine-grained semantic adaptation for industry-specific terminology. While existing deep learning models (such as BERT) can extract contextual semantics, their general pre-training mechanisms suffer from domain transfer bias in vertical skill disambiguation tasks. Furthermore, traditional attention mechanisms struggle to process local semantic interactions between skill phrases, limiting the alignment accuracy of cross-source skill data. These shortcomings make it difficult for existing systems to balance standardized accuracy and dynamic scalability when merging semantically diverse skill data from multiple sources.

[0004] Therefore, it is necessary to provide an optimized artificial intelligence-based employee skill map construction system to solve the above technical problems. Summary of the Invention

[0005] In order to solve the above technical problems, this application is proposed.

[0006] According to one aspect of the present application, a system for constructing an employee skills map based on artificial intelligence is provided, which includes: A multi-source skill data acquisition module is used to acquire multi-source skill data; A data preprocessing module is used to preprocess multi-source skill data to obtain structured skill data; A skill entity recognition module is used to perform skill entity recognition on structured skill data to obtain a list of skill phrases; A skill phrase mapping and disambiguation module is configured to perform skill phrase mapping and disambiguation on a list of skill phrases to obtain a list of standardized skill phrases, comprising: a skill phrase extraction unit configured to extract a first skill phrase and a second skill phrase from the list of skill phrases; a skill phrase mapping and matching unit configured to determine a fuzzy matching degree between the first skill phrase and the second skill phrase based on semantic-level co-coding features of the first skill phrase and the second skill phrase; and a skill phrase disambiguation unit configured to standardize the first skill phrase and the second skill phrase in response to the fuzzy matching degree being greater than or equal to a preset threshold to obtain a standard skill phrase corresponding to the first skill phrase and the second skill phrase. The skill knowledge graph construction module is used to build a skill knowledge graph based on a list of standardized skill phrases.

[0007] This application has at least the following technical effects: Compared with the existing technology, this application provides an artificial intelligence-based employee skill map construction system, which first obtains multi-source skill data, uses a deep learning-based data processing algorithm to process the multi-source skill data to mine skill phrases in the multi-source skill data, and further encodes the skill phrases into a high-dimensional semantic space, combines the semantic interaction mechanism to capture the semantic synergy relationship across skill descriptions, and realizes the semantic similarity judgment of skill phrases based on the semantic synergy relationship between skill phrases. By constructing a standardized skill system with strong semantic generalization capabilities, this application solves the shortcomings of traditional methods in eliminating semantic ambiguity in skill data, improves the accuracy of employee skill portrait construction, and provides accurate decision-making support for corporate talent strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 A block diagram of a system for building an employee skills map based on artificial intelligence according to an embodiment of the present application; Figure 2 A data flow diagram of a system for constructing an employee skills map based on artificial intelligence according to an embodiment of the present application; Figure 3 A block diagram of a skill phrase mapping and disambiguation module in an artificial intelligence-based employee skill map construction system according to an embodiment of the present application; Figure 4 This is a block diagram of a skill phrase mapping and matching unit in an artificial intelligence-based employee skill map construction system according to an embodiment of the present application. DETAILED DESCRIPTION

[0010] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0011] As used in this application and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not intended to refer to the singular but may include the plural. Generally speaking, the terms "comprises" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list. A method or apparatus may also include other steps or elements.

[0012] Although the present application makes various references to certain modules in the system according to embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are illustrative only, and different aspects of the system and method can use different modules.

[0013] Flowcharts are used in this application to illustrate the operations performed by the systems according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the various steps may be processed in reverse order or simultaneously, as needed. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0014] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0015] It should be noted that the acquisition and processing of all information or data in this application are carried out in compliance with the relevant national data protection laws and policies and with the authorization of the authority administrator.

[0016] In the technical solution of this application, an employee skill map construction system based on artificial intelligence is proposed. Figure 1 A block diagram of a system for building an employee skills map based on artificial intelligence according to an embodiment of the present application. Figure 2 This is a data flow diagram of the system for building an employee skills map based on artificial intelligence according to an embodiment of the present application. Figure 1 and Figure 2As shown, the employee skill map construction system 300 based on artificial intelligence according to an embodiment of the present application includes: a multi-source skill data acquisition module 310, used to acquire multi-source skill data; a data preprocessing module 320, used to perform data preprocessing on the multi-source skill data to obtain structured skill data; a skill entity recognition module 330, used to perform skill entity recognition on the structured skill data to obtain a list of skill phrases; a skill phrase mapping and disambiguation module 340, used to perform skill phrase mapping and disambiguation on the list of skill phrases to obtain a list of standardized skill phrases; a skill knowledge graph construction module 350, used to construct a skill knowledge graph based on the list of standardized skill phrases.

[0017] In particular, the multi-source skill data acquisition module 310 is used to acquire multi-source skill data. The multi-source skill data includes employee resume information, HR system data, performance evaluations, training records, project experience libraries, and internal skill certification data. It should be understood that employee resumes may only present basic skill labels, HR system data focuses on the static description of the job skill framework, while project experience libraries contain dynamic details of skill application (such as complex skills such as "e-commerce platform user profile construction"). Internal skill certification data also includes authoritative verification dimensions. Relying solely on a single source can easily lead to systematic deviations such as a lack of practical verification of skill profiles (such as the effectiveness of skill application in performance evaluations) and a break in skill evolution trajectories (such as skill iterations in training records). In the technical solution of this application, through the complementary fusion of cross-source data, a three-dimensional data base covering skill definition, application, verification, and evolution is constructed. This not only solves the problem of incomplete semantic coverage caused by data silos in traditional skill extraction, but also provides multi-perspective contextual support for subsequent semantic disambiguation. In particular, cross-verification of multi-source data can significantly improve the recall rate of skill entity recognition (such as mining the implicit skill "cross-departmental resource coordination ability" from performance reviews), while enhancing the credibility of skill association relationships (such as forming a skill advancement link between the internally certified "Python advanced development" and the "distributed system reconstruction" in the project). Ultimately, the constructed skill map has dynamic evolution characteristics, which can reflect the skill growth trajectory of employees in multiple scenarios such as training and improvement, project practice, etc. in real time, providing multi-dimensional and three-dimensional data penetration for corporate talent inventory.

[0018] In particular, the data preprocessing module 320 is used to perform data preprocessing on multi-source skill data to obtain structured skill data. It should be understood that in the process of constructing an employee skill map, the heterogeneity and noise interference of multi-source skill data are the core obstacles that hinder semantic analysis. In order to eliminate morphological noise and semantic interference in the original data, in the technical solution of the present application, data preprocessing is performed on multi-source skill data to obtain structured skill data, wherein the data preprocessing includes text extraction, structured conversion and data cleaning. Specifically, text extraction uses deep learning models (such as BiLSTM-CRF) to automatically identify skill keywords such as "Python programming" and "agile development" from unstructured data sources such as employee resumes and project documents. This addresses the difficulty of traditional manual annotation in covering implicit skill descriptions and ensures the integrity of skill information across data sources. Structured transformation uniformly maps semi-structured data such as HR system tables and training logs into a triplet format of (skill name, proficiency, source), eliminating processing bottlenecks caused by data format fragmentation. For example, "JAVA Development (Senior)" in performance evaluations and "Java Backend Engineer" in project experience can be converted into a unified dimension, providing structural consistency for subsequent semantic alignment. Data cleaning combines a rule engine with statistical models to filter out spelling variations in resumes (for example, correcting "Pyhon" to "Python"), merge duplicate entries caused by HR system input errors (for example, "machine learning" and "machine learning"), and identify outdated skill tags (for example, "Hadoop Ecosystem Development" needs to be associated with "Spark Computing Framework"), thereby improving the signal-to-noise ratio of the original data. These three layers of preprocessing mechanisms work together to integrate multimodal data, enabling implicit skills in unstructured text and explicit capabilities in system records to form computable data entities, providing high-quality input for high-precision semantic encoding and disambiguation, and ultimately supporting the dynamic evolution and precise reasoning of skill knowledge graphs.

[0019] In particular, the skill entity recognition module 330 is used to perform skill entity recognition on structured skill data to obtain a list of skill phrases, wherein skill phrases are phrases that describe employee skills. It should be understood that although the structured skill data has been unified in format, skill entities are still distributed in different data dimensions in an explicit or implicit manner - "leading data visualization projects" in performance evaluation may imply "Tableau chart design" skills, and "Kubernetes container orchestration" in project documents needs to be associated with the "cloud native architect" qualification in the certification system. In the technical solution of the present application, skill entity recognition is performed on structured skill data to obtain a list of skill phrases in the following steps: first, the structured skill data is segmented to obtain a list of skill data alternative words; then, skill entity recognition is performed on the list of skill data alternative words to obtain a list of skill entities; and then, a skill phrase extractor based on the TextRank model is used to extract skill phrases from the list of skill entities to obtain a list of skill phrases. That is, through a domain-optimized word segmentation algorithm, non-core semantic units are removed (such as filtering descriptive modifiers), and a context-aware entity recognition model is used to dynamically divide skill boundaries (for example, continuous semantic segments in long texts are identified as independent skill units). Key phrases are then extracted from co-occurrence relationships using a semantic weight model. By establishing a dynamic parsing mechanism for multi-source heterogeneous data, this not only addresses the blind spots in traditional rule matching for implicit skill descriptions (such as extracting unexplicitly labeled professional capabilities from project outcome descriptions), but also alleviates the problem of skill fragmentation caused by differences in expression habits (such as unifying the differentiated naming of the same capability by different departments into a standardized expression). The generated list of skill phrases not only covers basic capability labels, but also maps the dynamic evolution path of the skill system (such as the derivation of capabilities under emerging technology trends), providing high-quality input for subsequent high-dimensional semantic space mapping.

[0020] In particular, the skill phrase mapping and disambiguation module 340 is used to perform skill phrase mapping and disambiguation on the skill phrase list to obtain a list of standardized skill phrases. Figure 3 As shown, the skill phrase mapping and disambiguation module 340 includes: a skill phrase extraction unit 341, which is used to extract the first skill phrase and the second skill phrase from the list of skill phrases; a skill phrase mapping and matching unit 342, which is used to determine the fuzzy matching degree of the first skill phrase and the second skill phrase based on the semantic-level collaborative coding features of the first skill phrase and the second skill phrase; a skill phrase disambiguation unit 343, which is used to standardize the first skill phrase and the second skill phrase in response to the fuzzy matching degree being greater than or equal to a preset threshold to obtain a standard skill phrase corresponding to the first skill phrase and the second skill phrase.

[0021] Specifically, the skill phrase extraction unit 341 is configured to extract a first skill phrase and a second skill phrase from the list of skill phrases. It should be understood that multi-source data naturally exhibits semantic discreteness, and isolated processing of a single phrase cannot capture potential cross-expression associations. To overcome the limitations of traditional single-point disambiguation, a global semantic association topology is constructed to identify explicit synonymous relationships (such as standardized mappings of industry terms) in the list of skill phrases. In the technical solution of this application, first, the first skill phrase and the second skill phrase are extracted from the list of skill phrases. This allows the system to accurately capture the subtle differences and commonalities in skill descriptions by mapping each skill phrase into a high-dimensional semantic space and, on this basis, calculating fine-grained semantic co-coding features between the skill phrases. This ensures that the skill map not only contains rich skill information but also accurately reflects the actual skill level of employees.

[0022] Specifically, the skill phrase mapping matching unit 342 is used to determine the fuzzy matching degree between the first skill phrase and the second skill phrase based on the semantic level collaborative coding features of the first skill phrase and the second skill phrase. Figure 4 As shown, the skill phrase mapping and matching unit 342 includes: a semantic mapping subunit 3421, which is used to use the skill phrase embedding matrix to map the first skill phrase and the second skill phrase to the semantic latent space to obtain the high-dimensional semantic coding vector of the first skill phrase and the high-dimensional semantic coding vector of the second skill phrase; a semantic fine-grained interaction subunit 3422, which is used to input the high-dimensional semantic coding vector of the first skill phrase and the high-dimensional semantic coding vector of the second skill phrase into the semantic fine-grained interaction module to obtain the semantic fine-grained collaborative coding features between the first and second skill phrases; a fuzzy matching subunit 3423, which is used to obtain the fuzzy matching degree based on the semantic fine-grained collaborative coding features between the first and second skill phrases.

[0023] More specifically, the semantic mapping subunit 3421 is configured to use the skill phrase embedding matrix to map the first skill phrase and the second skill phrase into a semantic latent space, thereby obtaining a high-dimensional semantic encoding vector for the first skill phrase and a high-dimensional semantic encoding vector for the second skill phrase. It should be understood that the skill phrase list generated after structured processing of multi-source skill data still faces the challenge of semantic discreteness: the same core capability may exhibit varying granularity, industry terminology variations, or dynamic evolution across different data dimensions. The global semantic representation generated by traditional word embedding methods fails to capture the contextual sensitivity of skill descriptions (e.g., the semantic shift of a technical capability in different application scenarios) and the fine-grained differences in domain-specific semantics (e.g., the implicit boundary between general terminology and specialized terminology in specific fields). The pre-trained skill phrase embedding matrix enables mapping discrete textual representations into a high-dimensional semantic space, enabling a deep deconstruction of skill connotations at the latent feature level (e.g., mapping capabilities with different expression forms but consistent technical core into similar vector clusters). This also preserves the semantic topology driven by domain knowledge (e.g., the vector space distribution pattern of upstream and downstream relationships in the technology stack), providing context-rich feature input for subsequent semantic interaction modules.

[0024] More specifically, the semantic fine-grained interaction sub-unit 3422 is used to input the high-dimensional semantic encoding vector of the first skill phrase and the high-dimensional semantic encoding vector of the second skill phrase into the semantic fine-grained interaction module to obtain semantic fine-grained collaborative encoding features between the first and second skill phrases. It should be understood that the semantic complexity of skill descriptions is often reflected in the differences in the expression of the same skill in different data sources. For example, in the technical field, a core capability may be broken down into sub-skill descriptions at different levels or expressed using different combinations of terms due to industry practices, job characteristics, or personal expression habits of employees. When dealing with such scenarios, existing global semantic matching-based methods often fail to properly normalize semantically equivalent but heterogeneous skill phrases due to their difficulty in capturing the internal structural features of skill phrases and the local semantic associations across phrases. Therefore, to overcome the limitations of traditional single-vector similarity calculations and establish a fine-grained semantic mapping network across skill phrases, in the technical solution of this application, the high-dimensional semantic encoding vector of the first skill phrase and the high-dimensional semantic encoding vector of the second skill phrase are input into the semantic fine-grained interaction module to obtain semantic fine-grained collaborative encoding features between the first and second skill phrases. Specifically, the system first extracts local features from the high-dimensional semantic encoding vectors of skill phrases using one-dimensional convolution, enabling a slice-by-slice analysis of the semantic structure of skill phrases and identifying feature segments representing semantic components such as technical fields, operational objects, and application scenarios. Next, a single feature interaction engine matches and analyzes these local segments of skill phrases to identify local, fine-grained relationships between skill phrases. This local interaction mechanism effectively avoids misjudgments caused by structural differences in skill descriptions, often encountered in traditional methods. Furthermore, through the synergistic effect of a chained inference attention mechanism and LSTM, a fine-grained semantic collaborative encoding feature with hierarchical reasoning characteristics is generated between the first and second skill phrases. This significantly improves the system's adaptability to complex semantic scenarios. For example, when processing cross-industry skill data, the module can dynamically adjust the correlation strength between different industry terminologies using attention weights, completing semantic mapping while preserving domain specificity. In particular, the establishment of this fine-grained semantic collaboration capability provides a reliable semantic foundation for subsequent topological relationship reasoning in the skill knowledge graph, enabling enterprises to extract accurate skill profiles with industry and organizational characteristics from massive amounts of heterogeneous data.

[0025] Specifically, first, the high-dimensional semantic encoding vectors of the first skill phrase and the high-dimensional semantic encoding vectors of the second skill phrase are subjected to one-dimensional convolutional coding-based local latent semantic feature extraction to obtain a set of local semantic latent feature encoding vectors for the first skill phrase and a set of local semantic latent feature encoding vectors for the second skill phrase. It should be understood that the high-dimensional semantic encoding of skill descriptions often contains complex local semantic structures. For example, a skill phrase may be composed of multiple nested semantic components, such as the technology stack name, operation object, and application scenario. These components may appear in different orders or combinations in data from different sources. Traditional global vector similarity calculations compress the entire skill phrase into a single vector representation, which can easily flatten the hierarchical semantic information in the skill description and fail to distinguish local semantic differences such as "technical principles" and "technical applications." Therefore, in the technical solution of this application, the high-dimensional semantic encoding vectors of the first skill phrase and the high-dimensional semantic encoding vectors of the second skill phrase are subjected to one-dimensional convolutional coding-based local latent semantic feature extraction to obtain a set of local semantic latent feature encoding vectors for the first skill phrase and a set of local semantic latent feature encoding vectors for the second skill phrase. Here, local feature extraction effectively alleviates the problem of long-tail distribution of skill descriptions. For low-frequency but critical skill components (such as specific industry certification codes), the local perception characteristics of one-dimensional convolution can prevent them from being overwhelmed by high-frequency general vocabulary in the global vector, ensuring that unpopular but important skill features remain visible in subsequent interactive reasoning. This refined semantic dissection capability lays an interpretable semantic foundation for building a skill knowledge graph with industry characteristics. In a specific example of the present application, the following feature extraction formula is used to extract the local implicit semantic features of the skill phrases based on one-dimensional convolution coding for the high-dimensional semantic coding vector of the first skill phrase and the high-dimensional semantic coding vector of the second skill phrase to obtain a set of local semantic implicit feature coding vectors of the first skill phrase and a set of local semantic implicit feature coding vectors of the second skill phrase; wherein, the feature extraction formula is: in, is the high-dimensional semantic encoding vector of the first skill phrase, represents a one-dimensional convolution operation, are the first, second, and third in the set of local semantic implicit feature encoding vectors of the first skill phrase respectively. and The first skill phrase local semantic implicit feature encoding vector, is the high-dimensional semantic encoding vector of the second skill phrase, are the first, second, and third in the set of local semantic implicit feature encoding vectors of the second skill phrase respectively. and The local semantic implicit feature encoding vector of the second skill phrase.

[0026] Next, the set of local semantic implicit feature encoding vectors for the first skill phrase and each corresponding set of local semantic implicit feature encoding vectors for the second skill phrase in the set of local semantic implicit feature encoding vectors for the second skill phrase are respectively input into the skill phrase individual semantic feature interaction engine to obtain a set of first-second skill phrase local semantic implicit feature interaction response encoding vectors. It should be understood that after the high-dimensional semantic encoding vectors of the skill phrases undergo one-dimensional convolution decomposition, surface lexical noise is stripped away, forming local implicit semantic units such as "a certain algorithm optimization feature cluster" or "a certain engineering implementation pattern." However, the cross-phrase associations between these units are not explicitly modeled. Therefore, in the technical solution of the present application, the set of local semantic implicit feature encoding vectors for the first skill phrase and each corresponding set of local semantic implicit feature encoding vectors for the second skill phrase in the set of local semantic implicit feature encoding vectors for the first skill phrase are respectively input into the skill phrase individual semantic feature interaction engine to obtain a set of first-second skill phrase local semantic implicit feature interaction response encoding vectors. By pairing and interacting the local semantic units of skill phrases that have undergone one-dimensional convolution decomposition, the system can penetrate the surface semantic differences and capture the potential coupling relationships of local skill features such as "technical principle verification" and "production environment deployment". Through the adaptive learning of the fine-grained interaction engine, the system can automatically identify the implicit associations between skill phrases and form standardized skill nodes that are not predefined. This local interaction mechanism provides precise structured input for subsequent chain reasoning, so that the skill disambiguation process has the explainability of technical connotations, and ultimately supports enterprises to build dynamically evolving skill knowledge networks. In a specific example of the present application, the set of the first skill phrase local semantic implicit feature coding vectors and the set of the second skill phrase local semantic implicit feature coding vectors are input into the skill phrase monomer semantic feature interaction engine using the following feature interaction formula to obtain a set of first-second skill phrase local semantic implicit feature interaction response coding vectors; wherein, the feature interaction formula is: in, Indicates cascade, Indicates point multiplication by position, It means adding by position. Indicates difference by position. and denote the weight matrix and bias vector respectively, express and The corresponding first-second skill phrase local semantic implicit feature interaction response encoding vector.

[0027] Then, based on the semantic distribution characteristics of the local semantic implicit feature interaction response coding vectors of each first-second skill phrase in the set of local semantic implicit feature interaction response coding vectors of the first-second skill phrases, the chain reasoning attention weights of each first-second skill phrase local semantic implicit feature interaction response coding vector are determined to obtain a set of initial first-second skill phrase chain reasoning attention weights. It should be understood that the semantic equivalence between skill phrases is often determined by the interaction of several key local features, rather than a simple superposition of global semantics. Traditional methods use homogenized attention allocation and are unable to identify the dynamic correlation between local core semantic units such as "tuning" in "model tuning" and "optimization" in "parameter optimization", and are prone to misjudgment due to interference from non-key words. Therefore, in the technical solution of the present application, based on the semantic distribution characteristics of the local semantic implicit feature interaction response coding vectors of each first-second skill phrase in the set of local semantic implicit feature interaction response coding vectors of the first-second skill phrases, the chain reasoning attention weights of each first-second skill phrase local semantic implicit feature interaction response coding vector are determined to obtain a set of initial first-second skill phrase chain reasoning attention weights. That is, by quantifying the contribution of local interactions to semantic matching, the focusing and noise suppression of key semantic units are achieved. Here, by analyzing the semantic distribution characteristics of the local interaction response encoding vector (such as the spatial aggregation of the cosine similarity of the feature vector and the distribution offset represented by the KL divergence), the system can automatically identify local interaction nodes with domain knowledge relevance. This adaptive weight allocation mechanism based on feature distribution not only improves the fusion accuracy of cross-source skill data, but also provides topological guidance for semantic association paths for the dynamic expansion of the skill graph through the interpretable distribution of attention weights, so that the system can reason and extend along the existing semantic chain when facing emerging skill terms. In a specific example of the present application, the following attention formula is used to determine the chain reasoning attention weights of the local semantic implicit feature interaction response encoding vectors of each first-second skill phrase to obtain a set of initial first-second skill phrase chain reasoning attention weights, wherein the attention formula is: in, express function, The first and second skill phrases are represented by the set of local semantic implicit features of the interactive response encoding vectors. The feature value of the jth position of the local semantic implicit feature interaction response encoding vector of the first-second skill phrase, represents the square of the norm, represents the length of the first-second skill phrase local semantic implicit feature interaction response encoding vector, express Corresponding initial first-second skill phrase chain reasoning attention weights.

[0028] Then, the initial set of attention weights for the chained reasoning between the first and second skill phrases is spatially corrected based on the interaction norm to obtain the set of attention weights for the chained reasoning between the first and second skill phrases. In particular, it should be understood that when skill data from different sources (such as unstructured descriptions in employee resumes and standardized fields in HR systems) is mapped into a high-dimensional semantic space, the interaction patterns of their local implicit features exhibit two typical effects: one is a translational effect reflecting core semantic alignment, manifested as the consistency of feature gradient directions with the interaction paths; the other is a fluctuation effect arising from contextual differences or terminology variants, with feature gradient directions orthogonal to the interaction paths. Initial attention weights generated solely through conventional chained reasoning make it difficult to distinguish the impact of these two effects on the strength of semantic coupling. For example, when two skill phrases differ in their expression in a project experience database and training records, traditional methods may misjudge their association due to the local semantic perturbations caused by the fluctuation effect, resulting in skill entity splitting. Therefore, in a preferred embodiment of the present application, the initial set of attention weights for the chained reasoning between the first and second skill phrases is spatially corrected based on the interaction norm to obtain the set of attention weights for the chained reasoning between the first and second skill phrases. By deconstructing the translational and fluctuation components in the interaction space, the system can identify the stability and perturbation characteristics of the semantic relationships between skill phrases. For translational effects, the system strengthens the steady-state associations brought about by feature gradient consistency, ensuring stable alignment of core skill semantics. For fluctuation effects, a periodic compensation mechanism is used to suppress local semantic distortions caused by expression differences or industry terminology variations. For example, when processing heterogeneous skill descriptions from employees across departments, the system can automatically distinguish between terminology standardization differences (fluctuation effects) and core skill gaps (translation effects), thereby avoiding the inaccurate merging of different representations of the same skill into separate entities while accurately identifying substantial skill gaps. Spatially corrected attention weights dynamically balance the influence of semantic stability and local perturbations, making the skill disambiguation process both semantically sensitive and robust. This dynamic normative adaptation capability enables the constructed skill knowledge graph to achieve greater domain generalization, avoiding the rigid matching flaws of traditional rule engines while overcoming the transfer bias of general pre-trained models in specific verticals. Ultimately, this provides highly reliable data support for strategic decisions such as talent inventory and skills gap analysis.

[0029] As described above, when calculating the interaction response coding vectors of the local semantic implicit features of each first-second skill phrase, the interaction features between the corresponding local semantic implicit feature coding vectors of the first skill phrase and the local semantic implicit feature coding vectors of the second skill phrase are introduced, for example , , Etc., that is, they essentially correspond to different interaction norms, and thus have different spatial auxiliary constraint associations in the interaction space.

[0030] Therefore, in order to enhance the norm variability assistance in chain reasoning attention weights, it is preferred to modify the set of initial first-second skill phrase chain reasoning attention weights based on the spatial action decomposition of the interaction norm. Specifically, , is regarded as a translation effect, that is, the gradient direction is consistent with the spatial interaction direction, and It is regarded as a wave effect, that is, the gradient direction is orthogonal to the spatial interaction direction. In this way, for vector statistics of different effects, for example , , , it is believed that the wave action leads to a localized periodic effect in the translation direction, and thus the periodic local regular compensation is calculated as: in, and represents the translation action vector statistics, represents the statistics of the wave action vector, represents periodic local regular compensation; That is, As a localized representation of translation, the wave action It increases with the logarithmic size of the translation domain.

[0031] On the other hand, the wave action will also lead to a localized translational phase transition, thus obtaining a periodic local auxiliary phase transition: in, Represents the periodic local auxiliary phase transition; thus, the initial first-second skill phrase chain reasoning attention weight is modified based on the weighted sum of the above two items : By distinguishing the normative roles of different interaction norms corresponding to the local interaction responses of each first-second skill phrase in the interaction space, the correlation between different spatial auxiliary constraint norms in the interaction space can be improved, thereby improving the calculation accuracy of the chain reasoning attention weight.

[0032] Furthermore, based on the set of attention weights of the first-second skill phrase chain reasoning, the set of local semantic implicit feature interaction response coding vectors of the first-second skill phrase is weighted modulated to obtain a set of modulated first-second skill phrase local semantic implicit feature interaction response coding vectors. It should be understood that the semantic coupling pattern of multi-source heterogeneous skill data has significant non-uniformity characteristics. When the attention weights of the first-second skill phrase chain reasoning are generated by local interaction chain reasoning, the implicit feature interaction process will simultaneously include core semantic alignment signals and redundant noise interference. The traditional attention mechanism cannot adaptively distinguish between key related features and background noise in such mixed interaction patterns, resulting in the semantic similarity calculation being susceptible to local fluctuation interference. Therefore, in the technical solution of the present application, based on the set of attention weights of the first-second skill phrase chain reasoning, the set of local semantic implicit feature interaction response coding vectors of the first-second skill phrase is weighted modulated to obtain a set of modulated first-second skill phrase local semantic implicit feature interaction response coding vectors. Specifically, the system leverages the spatial distribution of chained inference attention weights to achieve hierarchical extraction of interaction information. Specifically, the system analyzes the clustering patterns of chained inference attention weights between the first and second skill phrases in the interaction space, identifying high-response regions reflecting the essential semantic connections between skill phrases and weighting the corresponding local feature vectors. Conversely, feature groups with discrete weight distributions and low response strength are identified as local perturbations caused by differences in representation or industry terminology variations, and their impact on the overall similarity metric is reduced through weight suppression. Through dynamic weight allocation, the system effectively captures the hierarchical semantic structure implicit in multi-source skill descriptions. In the core skill dimension, it strengthens the steady-state semantic associations across representations to ensure the accurate merging of synonymous skill phrases. In the sub-domain dimension, it mitigates local interference signals caused by data source differences or contextual shifts to avoid overfitting during semantic disambiguation. This feature selection capability enables the constructed skill knowledge graph to maintain the rigor of the industry terminology system while maintaining the flexibility to adapt to dynamic business expansion, providing a highly reliable semantic analysis foundation for enterprise talent development and skill upgrade path planning. In a specific example of the present application, the set of first-second skill phrase local semantic implicit feature interaction response coding vectors is weighted modulated using the following modulation formula to obtain a set of modulated first-second skill phrase local semantic implicit feature interaction response coding vectors; wherein the modulation formula is: in, express The corresponding first-second skill phrase chain reasoning attention weight, is a set of the local semantic implicit feature interaction response encoding vectors of the modulation first-second skill phrase, are the first, second, and third in the set of the local semantic implicit feature interaction response encoding vectors of the modulation first-second skill phrases. and Modulate the first-second skill phrase local semantic implicit feature interaction response encoding vector.

[0033] Subsequently, the set of the interaction response coding vectors of the local semantic implicit features of the modulated first and second skill phrases is input into the skill phrase chain reasoning engine based on the forward LSTM model to obtain the semantic fine-grained collaborative coding vector between the first and second skill phrases as the semantic fine-grained collaborative coding feature between the first and second skill phrases. It should be understood that the set of the interaction response coding vectors of the local semantic implicit features of the modulated first and second skill phrases essentially constitutes a semantic reasoning chain with potential causal associations. Traditional models find it difficult to capture long-span dependencies between local features. For example, the implicit association of a skill phrase in project experience may need to span multiple local interaction units to be fully presented. Therefore, in the technical solution of the present application, the set of the interaction response coding vectors of the local semantic implicit features of the modulated first and second skill phrases is input into the skill phrase chain reasoning engine based on the forward LSTM model to obtain the semantic fine-grained collaborative coding vector between the first and second skill phrases. Specifically, the forward LSTM model regards the set of the interaction response coding vectors of the local semantic implicit features of the modulated first and second skill phrases as an ordered sequence of semantic reasoning units, and uses its gating mechanism to gradually accumulate context state information. For example, when processing complex skill descriptions in cross-departmental collaboration scenarios, the LSTM memory unit can continuously track the semantic evolution of core skill dimensions while dynamically filtering out temporary representational noise caused by data source differences through a forget gate. This sequential reasoning mechanism enables the system to identify implicit, progressive semantic connections between skill phrases, such as the upgrade path of competency requirements for a specific skill across different project phases. The resulting fine-grained semantic co-encoding vector between the first and second skill phrases comprehensively reflects the multi-level and multi-dimensional strength of semantic connections between skill phrases. For example, when integrating an R&D team's technical stack description with a customer requirements document, the system can use chained reasoning to capture the semantic extension of technical terminology from fundamental principles to application scenarios, accurately identifying the equivalence between variant expressions of the same skill in different contexts and its core competencies. This global semantic representation not only overcomes the limitations of traditional attention mechanisms in modeling long-span dependencies but also enhances the system's adaptability to the dynamic evolution of industry terminology, such as the changing names of emerging technologies. The resulting skill knowledge graph exhibits stronger semantic coherence and business interpretability, providing reliable technical support for enterprises to build dynamic talent competency models and accurately identify skill transformation needs. In a specific example of the present application, the set of the modulated first-second skill phrase local semantic implicit feature interaction response encoding vectors is input into the skill phrase chain inference engine based on the forward LSTM model to obtain the semantic fine-grained collaborative encoding vector between the first and second skill phrases; wherein the chain inference formula is: in, Indicates forward , It is the semantic fine-grained collaborative encoding vector between the first and second skill phrases.

[0034] More specifically, the fuzzy matching subunit 3423 is configured to obtain a fuzzy matching degree based on the semantic fine-grained co-encoding features between the first and second skill phrases. In the technical solution of this application, the semantic fine-grained co-encoding vector between the first and second skill phrases is input into a decoder-based semantic fuzziness measurement module to obtain a fuzzy matching degree. It should be understood that when enterprises integrate skill descriptions from multiple channels, such as recruitment systems, project documents, and performance evaluations, the same core competency often exhibits diverse expressions due to differences in departmental terminology or contextual shifts. Traditional similarity calculation methods (such as cosine similarity) can capture shallow semantic associations, but they cannot handle complex semantic relationships between skill phrases that are "partially overlapping but not identical." For example, the semantic inclusiveness and hierarchical differences between the deep application of a technical concept in R&D scenarios and its basic operational guidance in after-sales support exist. In such scenarios, relying solely on global semantic matching can easily lead to skill entity merging errors, affecting the accuracy of graph construction. In the technical solution of this application, to construct a dynamic and interpretable semantic association quantification mechanism, the nonlinear mapping capabilities of the decoder are used to deconstruct the implicit semantic coupling patterns between skill phrases. The semantic fuzziness measurement module, using a decoder network, converts the semantically fine-grained co-encoding vectors between the first and second skill phrases into a business-interpretable matching score. Its design fully considers the multi-layered nature of industry knowledge systems. For example, when processing complex skill descriptions generated through cross-departmental collaboration, the module distinguishes between strongly correlated features of the core skill dimension (e.g., key implementation steps in a professional methodology) and weakly correlated features of the auxiliary skill dimensions (e.g., differences in operational procedures across different business scenarios). Using a hierarchical decoding strategy, it generates a fuzzy matching score that reflects the strength of semantic equivalence. This quantification mechanism enables the system to identify semantically equivalent skill pairs with heterogeneous representations while effectively eliminating distracting items that are superficially similar but essentially unrelated. Leveraging the decoder's feature decoupling capabilities, the semantic fuzziness measurement module accurately captures complex relationships between skill phrases, such as "asymmetric inclusion" and "partial overlap," providing a refined basis for subsequent standardized annotation decisions. This dynamic semantic quantification capability ensures that the constructed skill knowledge graph combines standardized accuracy with business scalability, providing reliable data support for multi-dimensional analysis of enterprise talent inventories (e.g., identifying skill gaps and assessing cross-job fit), effectively overcoming the technical bottlenecks of traditional methods in semantic disambiguation and entity alignment.

[0035] Specifically, the skill phrase disambiguation unit 343 is configured to, in response to a fuzzy matching degree being greater than or equal to a preset threshold, standardize the first skill phrase and the second skill phrase to obtain a standard skill phrase corresponding to the first skill phrase and the second skill phrase. That is, in the technical solution of the present application, in response to a fuzzy matching degree being greater than or equal to a preset threshold, determine that the first skill phrase and the second skill phrase are a synonymous phrase pair, and standardize the first skill phrase and the second skill phrase based on a skill phrase dictionary to obtain a standard skill phrase corresponding to the first skill phrase and the second skill phrase; wherein the standard skill phrase is used to replace the first skill phrase and the second skill phrase. It should be understood that traditional matching mechanisms based on fixed rules have difficulty distinguishing such complex relationships between semantic inclusiveness and representational heterogeneity, resulting in the standardization process being prone to over-merging or erroneous splitting. Therefore, in the technical solution of the present application, in order to construct a dynamically interpretable semantic equivalence decision mechanism, a balance is achieved between standardization accuracy and domain extensibility through adaptive adjustment of preset thresholds. Specifically, when the fuzzy match exceeds a threshold, the system determines that the two skill phrases have a sufficiently strong semantic synergy, triggering a standardized tagging process for synonymous phrase pairs. For example, when processing multilingual skill data for a cross-border M&A enterprise, the system can automatically identify semantic equivalence between skill expressions in different languages (such as the mapping between English technical certification names and localized job competency descriptions) through a threshold mechanism, while also avoiding the incorrect merging of industry-wide basic skills with advanced skills in specific fields. This standardized mechanism ensures the constructed skill knowledge graph is both authoritative and provides multi-granular data support for enterprises to implement precise talent inventory and develop personalized training plans.

[0036] In particular, the skill knowledge graph construction module 350 is used to construct a skill knowledge graph based on a list of standardized skill phrases. Specifically, in a specific example of the present application, based on the list of standardized skill phrases, the system embeds discrete skill nodes into the industry knowledge system through ontology modeling technology, and constructs a three-dimensional semantic network including a skill classification tree, capability hierarchy, and association relationships. For example, when processing the technology stack evolution data of the R&D department, the system automatically identifies the hierarchical relationship between core basic skills and derivative technologies, captures the cross-fusion nodes of emerging technology fields and traditional technology systems, and forms a skill development map with a time dimension. At the same time, combined with the skill co-occurrence pattern in the employee project experience, the implicit skill combination rules (such as the high-frequency collaborative appearance of a group of skills in innovative projects) are excavated to provide data support for team capability optimization. That is, the hierarchical organization and intelligent reasoning of the skill system are realized through the multi-dimensional relationship network of the knowledge graph. It's worth noting that the constructed skills knowledge graph not only visualizes skill data but also unlocks data value through a relational reasoning engine. In talent inventory scenarios, the graph's path analysis function automatically derives the optimal path for employee skill upgrades, generating personalized development recommendations based on historical training records and job requirements. At the organizational planning level, skill co-occurrence pattern mining can identify key capability gaps in corporate strategic transformation, providing a decision-making basis for cross-departmental talent deployment. This dynamic knowledge system enables companies to shift their talent strategies from experience-driven to data-driven, building intelligent organizational capabilities adapted to the demands of the digital economy.

[0037] It should be noted that those skilled in the art will recognize that, in addition to the aforementioned technical solution of "building a skill knowledge graph based on a list of standardized skill phrases," other existing technologies can also be employed to implement this technical process. For example, in another specific implementation, the skill knowledge graph construction solution in the paper "SkillKG: A Knowledge Graph for Skill-Driven" can be used to construct the knowledge graph envisioned by Career Path Recommendation. It should be understood that "SkillKG: A Knowledge Graph for Skill-Driven Career Path Recommendation" is prior art, and to avoid elaboration, its content will not be elaborated upon here.

[0038] As described above, the employee skill map construction system 300 based on artificial intelligence according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with an employee skill map construction algorithm based on artificial intelligence. In one possible implementation, the employee skill map construction system 300 based on artificial intelligence according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the employee skill map construction system 300 based on artificial intelligence can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the employee skill map construction system 300 based on artificial intelligence can also be one of the many hardware modules of the wireless terminal.

[0039] Alternatively, in another example, the artificial intelligence-based employee skill map construction system 300 and the wireless terminal may also be separate devices, and the artificial intelligence-based employee skill map construction system 300 may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0040] While various embodiments of the present disclosure have been described above, the above descriptions are illustrative, non-exhaustive, and not intended to be limiting of the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or improvements to existing technologies, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An employee skills map construction system based on artificial intelligence, characterized by: include: A multi-source skill data acquisition module is used to acquire multi-source skill data; A data preprocessing module is used to preprocess multi-source skill data to obtain structured skill data; A skill entity recognition module is used to perform skill entity recognition on structured skill data to obtain a list of skill phrases; A skill phrase mapping and disambiguation module is configured to perform skill phrase mapping and disambiguation on a list of skill phrases to obtain a list of standardized skill phrases, comprising: a skill phrase extraction unit configured to extract a first skill phrase and a second skill phrase from the list of skill phrases; a skill phrase mapping and matching unit configured to determine a fuzzy matching degree between the first skill phrase and the second skill phrase based on semantic-level co-coding features of the first skill phrase and the second skill phrase; and a skill phrase disambiguation unit configured to standardize the first skill phrase and the second skill phrase in response to the fuzzy matching degree being greater than or equal to a preset threshold to obtain a standard skill phrase corresponding to the first skill phrase and the second skill phrase. The skill knowledge graph construction module is used to build a skill knowledge graph based on a list of standardized skill phrases.

2. The employee skill map construction system based on artificial intelligence according to claim 1 is characterized in that: Multi-source skills data includes employee resume information, HR system data, performance evaluations, training records, project experience libraries, and internal skills certification data.

3. The employee skill map construction system based on artificial intelligence according to claim 2 is characterized in that: Data preprocessing includes text extraction, structured transformation and data cleaning.

4. The employee skill map construction system based on artificial intelligence according to claim 3 is characterized in that: Skill entity recognition module, used to: Segmenting the structured skill data to obtain a list of candidate words for the skill data; Performing skill entity recognition on the list of skill data candidate words to obtain a list of skill entities; A skill phrase extractor based on the TextRank model is used to extract skill phrases from the list of skill entities to obtain a list of skill phrases.

5. The employee skill map construction system based on artificial intelligence according to claim 4 is characterized in that: Skill phrase mapping matching unit, including: a semantic mapping subunit, configured to map the first skill phrase and the second skill phrase into a semantic latent space using a skill phrase embedding matrix to obtain a high-dimensional semantic encoding vector of the first skill phrase and a high-dimensional semantic encoding vector of the second skill phrase; A semantic fine-grained interaction subunit, configured to input the high-dimensional semantic encoding vector of the first skill phrase and the high-dimensional semantic encoding vector of the second skill phrase into a semantic fine-grained interaction module to obtain semantic fine-grained collaborative encoding features between the first and second skill phrases; The fuzzy matching subunit is used to obtain the fuzzy matching degree based on the semantic fine-grained collaborative coding features between the first and second skill phrases.

6. The employee skill map construction system based on artificial intelligence according to claim 5 is characterized in that: Semantic fine-grained interaction sub-units, including: A skill phrase local semantic interaction secondary subunit is used to calculate a set of first-second skill phrase local semantic implicit feature interaction response encoding vectors based on the skill phrase local semantic implicit encoding features of the first skill phrase high-dimensional semantic encoding vector and the second skill phrase high-dimensional semantic encoding vector; a feature modulation secondary subunit, configured to perform feature modulation based on a chained reasoning attention mechanism on the set of first-second skill phrase local semantic implicit feature interaction response encoding vectors based on the semantic distribution characteristics of each first-second skill phrase local semantic implicit feature interaction response encoding vector in the set of first-second skill phrase local semantic implicit feature interaction response encoding vectors, so as to obtain a set of modulated first-second skill phrase local semantic implicit feature interaction response encoding vectors; The chain reasoning secondary sub-unit is used to input the set of interactive response coding vectors of the local semantic implicit features of the modulated first and second skill phrases into the skill phrase chain reasoning engine based on the forward LSTM model to obtain the semantic fine-grained collaborative coding vector between the first and second skill phrases as the semantic fine-grained collaborative coding feature between the first and second skill phrases.

7. The employee skill map construction system based on artificial intelligence according to claim 6 is characterized in that: Skill phrase local semantic interaction secondary sub-unit, used for: Performing skill phrase local latent semantic feature extraction based on one-dimensional convolutional coding on the high-dimensional semantic coding vector of the first skill phrase and the high-dimensional semantic coding vector of the second skill phrase respectively to obtain a set of local semantic latent feature coding vectors of the first skill phrase and a set of local semantic latent feature coding vectors of the second skill phrase; The set of local semantic implicit feature coding vectors of the first skill phrase and the set of local semantic implicit feature coding vectors of the second skill phrase corresponding to each group of local semantic implicit feature coding vectors of the first skill phrase and the set of local semantic implicit feature coding vectors of the second skill phrase are respectively input into the skill phrase monomer semantic feature interaction engine to obtain a set of first-second skill phrase local semantic implicit feature interaction response coding vectors.

8. The employee skill map construction system based on artificial intelligence according to claim 7 is characterized in that: Characteristic modulation secondary subunit, used for: Determining, based on the semantic distribution characteristics of each first-second skill phrase local semantic implicit feature interaction response encoding vector in the set of first-second skill phrase local semantic implicit feature interaction response encoding vectors, a chain reasoning attention weight of each first-second skill phrase local semantic implicit feature interaction response encoding vector to obtain an initial first-second skill phrase chain reasoning attention weight set; Performing a spatial correction based on the interaction norm on the set of initial first-second skill phrase chain reasoning attention weights to obtain a set of first-second skill phrase chain reasoning attention weights; Based on the set of attention weights of the first-second skill phrase chain reasoning, the set of local semantic implicit feature interaction response encoding vectors of the first-second skill phrase is weighted modulated to obtain a set of modulated first-second skill phrase local semantic implicit feature interaction response encoding vectors.

9. The employee skill map construction system based on artificial intelligence according to claim 8 is characterized in that: Fuzzy matching subunit, used for: The semantic fine-grained co-encoding vector between the first and second skill phrases is input into the decoder-based semantic fuzzy measurement module to obtain the fuzzy matching degree.

10. The employee skill map construction system based on artificial intelligence according to claim 9 is characterized in that: Skill phrase disambiguation unit, used to: In response to the fuzzy matching degree being greater than or equal to a preset threshold, determining that the first skill phrase and the second skill phrase are a synonym phrase pair; The first skill phrase and the second skill phrase are normalized based on the skill phrase dictionary to obtain standard skill phrases corresponding to the first skill phrase and the second skill phrase.

Images