A method, system and medium for extracting content features of a personnel configuration CMS

Abstract

The present invention discloses a content feature extraction method, system and medium for a personnel configuration CMS, and relates to the technical field of content feature extraction. The method generates a sentence feature vector for each sentence text, uses a pre-trained language model to extract context feature representation of text content data, constructs a semantic relationship graph between sentences based on the sentence feature vector and the context feature representation, collects a personnel configuration domain knowledge base, expands the semantic relationship graph, uses a feature fusion strategy, generates a fused sentence feature vector through the expanded semantic relationship graph, clusters the fused sentence feature vectors, obtains K semantic clustering clusters, extracts keywords representing the semantic clustering cluster from each semantic clustering cluster, automatically integrates context semantics and structured knowledge, and improves the migration and generalization capabilities of domain knowledge, providing strong support for intelligent content analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of content feature extraction, and in particular to a method, system and medium for extracting content features of a personnel configuration CMS. Background Art

[0002] Today, the personnel management of enterprises pays more and more attention to digitalization and intelligence. As a key system for centralized storage and management of a large amount of unstructured text materials such as employee manuals, training materials, recruitment information, etc., the personnel content management system is facing huge challenges in content semantic mining and knowledge modeling. These massive text data contain rich personnel management knowledge, such as job responsibilities, technical requirements, salary and benefits, career development plans, etc., which are crucial for enterprises to make personnel strategic plans, optimize talent training programs, and improve employee experience. However, due to the unstructured nature of text content, traditional rule-based or keyword-based analysis methods are difficult to effectively mine its deep semantic information.

[0003] The Chinese patent with the authorization announcement number CN116644184B discloses a human resources information management system based on data clustering, including a data acquisition module for acquiring keyword areas of each dimension of human resources information and setting corresponding dimension weights; a comprehensive advantage analysis module for acquiring non-repeating texts in the keyword area and acquiring comprehensive advantage parameters of each dimension in each piece of human resources information; a dimension advantage analysis module for acquiring the same-dimensional advantage parameters of each dimension according to the hierarchical weights of the same dimension of human resources information; and clustering management of human resources information by combining the comprehensive advantage parameters, same-dimensional advantage parameters and dimension weights of any dimension in each piece of human resources information. This solution fails to solve the problem of parsing the contextual association relationship in human resources texts;

[0004] To this end, the present invention proposes a method, system and medium for extracting content features of a personnel configuration CMS. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a method, system and medium for extracting content features of a personnel configuration CMS, which automatically integrates contextual semantics and structured knowledge, and improves the transfer and generalization capabilities of domain knowledge, providing strong support for intelligent content analysis.

[0006] To achieve the above purpose, a method for extracting content features of a staffing CMS is proposed, which includes the following steps:

[0007] Step 1: Collect text content data of personnel configuration CMS;

[0008] Step 2: Segment the text content data into sentences to obtain a set of sentence texts, segment each sentence text, convert each segmented word into a word vector, and generate a sentence feature vector for each sentence text;

[0009] Step 3: Use the pre-trained language model to extract the contextual feature representation of the text content data, and build a semantic relationship graph between sentences based on the sentence feature vector and the contextual feature representation;

[0010] Step 4: Collect the knowledge base of staffing domain and expand the semantic relationship graph based on the knowledge of staffing domain;

[0011] Step 5: Using the feature fusion strategy, the sentence feature vector and context feature representation of each sentence text are fused through the expanded semantic relationship graph to generate a fused sentence feature vector;

[0012] Step 6: Cluster the fused sentence feature vector to obtain K semantic clusters; K is the number of semantic clusters obtained;

[0013] Step 7: Extract keywords representing each semantic cluster from each semantic cluster;

[0014] The collecting of text content data of personnel configuring CMS comprises the following steps:

[0015] Step 11: Identify the staffing CMS data sources and collect text data from the staffing CMS data sources;

[0016] Step 12: Clean and annotate the text data to obtain text content data;

[0017] The method of segmenting the text content data into sentences and obtaining a sentence text set is as follows:

[0018] Based on language rules, the text content data is subjected to sentence boundary recognition, and the text content data is divided into different sentence texts according to the sentence boundaries to form a sentence text set;

[0019] The method of segmenting each sentence text, converting each segmented word into a word vector, and generating a sentence feature vector for each sentence text is as follows:

[0020] Use the word segmentation tool to divide each sentence text into words and obtain the word set corresponding to each sentence text;

[0021] Use the pre-trained word vector model to convert each word in the word set into a word vector form;

[0022] The word vectors of each word in the word set corresponding to each sentence text are sequentially combined into a word vector set;

[0023] The word vector set is converted into a comprehensive feature vector of a sentence using a sequence model, where the comprehensive feature vector of the sentence is the sentence feature vector;

[0024] The method of extracting context feature representation of text content data using a pre-trained language model comprises the following steps:

[0025] Step 21: input each sentence text into a pre-trained language model, encode the sentence text, and obtain a contextual semantic representation vector of each word in the sentence text; the pre-trained language model is a Bert model;

[0026] Step 22: Input the sequence of contextual semantic representation vectors of the word units of each sentence text into the Hierarchical Attention Network model to obtain a sentence-level context vector representation, which is the context feature representation;

[0027] The method of constructing a semantic relationship graph between sentences based on sentence feature vectors and context feature representations includes the following steps:

[0028] Step 31: concatenate the sentence feature vector and context vector of each sentence text to form a sentence global feature vector of the sentence text;

[0029] Step 32: For the sentence global feature vectors corresponding to any two sentence texts, calculate the cosine similarity of the two sentence global feature vectors. If the cosine similarity is greater than a preset similarity threshold, the two sentence texts are regarded as having a semantically related edge.

[0030] Step 33: Take each sentence as a sentence node, and connect two sentence nodes with semantically associated edges using a weighted edge; the weight of the weighted edge is the cosine similarity corresponding to the semantically associated edge;

[0031] Step 34: All sentence nodes and weighted edges between sentence nodes form a semantic relationship graph;

[0032] The collecting of the staffing domain knowledge base and the expansion of the semantic relationship graph based on the staffing domain knowledge include the following steps:

[0033] Step 41: Collect structured data from the field of staffing; the structured data includes but is not limited to job descriptions, skill requirements, salary information, etc.;

[0034] Step 42: Extract job triple relationships from structured data; import the extracted triple relationships into a knowledge graph tool, construct an ontology hierarchy and relationship, and obtain a domain knowledge graph;

[0035] Step 43: Use a named entity recognition algorithm to perform named entity recognition on each sentence text in the text content data to obtain a target entity set;

[0036] Step 44: Find the connection path between the target entity sets of any two sentence texts in the domain knowledge graph, and based on the connection path, construct an extended weighted edge between every two sentence nodes in the semantic relationship graph, where the weight of the extended weighted edge is the inverse of the length of the connection path;

[0037] The feature fusion strategy is used to fuse the sentence feature vector and context feature representation of each sentence text through the expanded semantic relationship graph to generate the fused sentence feature vector in the following manner:

[0038] Inputting the expanded semantic relationship graph into a graph attention network, the graph attention network automatically learns the semantic relationship weight between every two sentence nodes;

[0039] For each sentence node, according to its relationship weight with other nodes and the sentence global feature vector of the sentence node, the graph attention network generates a new sentence vector representation that integrates context and knowledge. The new sentence vector representation is the fused sentence feature vector.

[0040] The method of extracting the keywords representing each semantic cluster from each semantic cluster is as follows:

[0041] For each semantic cluster, calculate the average cosine similarity between its cluster centroid and the fused sentence feature vector of all the sentence texts in it;

[0042] The K1 semantic clusters with the highest average cosine similarity are taken as key clusters; K1 is a preset quantity parameter;

[0043] In each key cluster, the words are sorted according to the similarity score between the contextual semantic representation vector of each word and the centroid of the key cluster, and the first N1 words are used as the keywords of the key cluster.

[0044] A content feature extraction system for staffing CMS is proposed, which includes a text data collection module, a semantic relationship graph construction module, a feature fusion module and a keyword extraction module; wherein each module is electrically connected;

[0045] The text data collection module collects the text content data of the personnel configuring the CMS and sends the text content data to the semantic relationship graph construction module;

[0046] The semantic relationship graph construction module divides the text content data into sentences, obtains a sentence text set, performs word segmentation on each sentence text, converts each segmented word into a word vector, generates a sentence feature vector for each sentence text, extracts the context feature representation of the text content data using a pre-trained language model, constructs a semantic relationship graph between sentences based on the sentence feature vector and the context feature representation, collects a knowledge base in the field of staffing, expands the semantic relationship graph based on the knowledge in the field of staffing, and sends the semantic relationship graph to the feature fusion module;

[0047] The feature fusion module uses the feature fusion strategy to fuse the sentence feature vectors through the semantic relationship graph to generate a fused sentence feature vector, and sends the fused sentence feature vector of each sentence text to the keyword extraction module;

[0048] The keyword extraction module clusters the fused sentence feature vectors to obtain K semantic clusters, and extracts keywords representing each semantic cluster from each semantic cluster.

[0049] A computer-readable storage medium is provided, on which a rewritable computer program is stored;

[0050] When the computer program is executed on a computer device, the computer device is enabled to execute the above-mentioned method for extracting content features of a personnel configuration CMS.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The present invention collects text content data of a personnel configuration CMS, performs sentence segmentation on the text content data, obtains a sentence text set, performs word segmentation on each sentence text, performs word vector conversion on each segmented word, generates a sentence feature vector for each sentence text, uses a pre-trained language model to extract context feature representation of the text content data, constructs a semantic relationship graph between sentences based on the sentence feature vector and the context feature representation, collects a personnel configuration domain knowledge base, and expands the semantic relationship graph based on the personnel configuration domain knowledge, uses a feature fusion strategy, fuses the sentence feature vector and the context feature representation of each sentence text through the expanded semantic relationship graph, generates a fused sentence feature vector, clusters the fused sentence feature vector, obtains K semantic clustering clusters, and extracts keywords representing the semantic clustering cluster from each semantic clustering cluster; the scheme captures context semantics through a pre-trained language model, fuses the personnel configuration domain knowledge graph, and introduces a graph attention network to adaptively model the relationship between different semantic units, thereby generating a rich semantic representation that integrates context and knowledge. On this basis, the semantic vectors are further clustered to automatically discover key semantic topic clusters in the text content and provide interpretable centroid vectors to present the core semantics. The contextual semantics and structured knowledge are automatically integrated, and the transfer and generalization capabilities of domain knowledge are improved, providing strong support for intelligent content analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A flowchart of a method for extracting content features of a personnel configuration CMS in Embodiment 1 of the present invention;

[0054] Figure 2 A module connection relationship diagram of a personnel configuration CMS content feature extraction system in Example 2 of the present invention;

[0055] Figure 3 This is a schematic diagram of the computer-readable storage medium structure in Example 3 of the present invention. DETAILED DESCRIPTION

[0056] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] Example 1

[0058] like Figure 1 As shown, a method for extracting content features of a personnel configuration CMS includes the following steps:

[0059] Step 1: Collect text content data of personnel configuration CMS;

[0060] Step 2: Segment the text content data into sentences to obtain a set of sentence texts, segment each sentence text, convert each segmented word into a word vector, and generate a sentence feature vector for each sentence text;

[0061] Step 3: Use the pre-trained language model to extract the contextual feature representation of the text content data, and build a semantic relationship graph between sentences based on the sentence feature vector and the contextual feature representation;

[0062] Step 4: Collect the knowledge base of staffing domain and expand the semantic relationship graph based on the knowledge of staffing domain;

[0063] Step 5: Using the feature fusion strategy, the sentence feature vector and context feature representation of each sentence text are fused through the expanded semantic relationship graph to generate a fused sentence feature vector;

[0064] Step 6: Cluster the fused sentence feature vector to obtain K semantic clusters; K is the number of semantic clusters obtained;

[0065] Step 7: Extract keywords representing each semantic cluster from each semantic cluster;

[0066] The step of collecting text content data of the CMS configured by the personnel includes the following steps:

[0067] Step 11: Identify the staffing CMS data sources and collect text data from the staffing CMS data sources;

[0068] Specifically, the data sources of the staffing CMS include but are not limited to the internal staffing management system of the enterprise, the staffing-related documents issued by the government and institutions, online recruitment websites and job description libraries, etc.; the text data can be collected by web crawler technology, database reading or exporting tools;

[0069] Step 12: Clean and annotate the text data to obtain text content data;

[0070] Specifically, the method of cleaning and labeling text data may be:

[0071] Use pattern matching method to remove HTML tags, special characters, junk data, etc. from text data;

[0072] Use duplicate checking tools to detect and remove duplicate data in text data;

[0073] Manually or automatically annotate the data with metadata, including but not limited to data source, release time, industry field, document type, etc.;

[0074] Furthermore, the text content data is segmented into sentences to obtain a sentence text set in the following manner:

[0075] Based on language rules, the text content data is subjected to sentence boundary recognition, and the text content data is divided into different sentence texts according to the sentence boundaries to form a sentence text set;

[0076] Specifically, the language rule may be a rule based on language symbols, such as regarding the text between two periods as a sentence;

[0077] Furthermore, the method of segmenting each sentence text, converting each segmented word into a word vector, and generating a sentence feature vector for each sentence text is as follows:

[0078] Use a word segmentation tool to segment each sentence text to obtain a word set corresponding to each sentence text; the word segmentation tool may be jieba, SnowNLP, etc.;

[0079] Using a pre-trained word vector model to convert each word in the word set into a word vector form; the pre-trained word vector model may be a word2vec model or a GloVe model, etc.;

[0080] The word vectors of each word in the word set corresponding to each sentence text are sequentially combined into a word vector set;

[0081] Using a sequence model to convert a word vector set into a comprehensive feature vector of a sentence, the comprehensive feature vector of the sentence is a sentence feature vector; the sequence model may be an RNN model or a Transformer model;

[0082] Specifically, when the sequence model is an RNN model:

[0083] The word vector set is used as the input of the RNN model, and the context information is obtained through the hidden layer loop;

[0084] Use the hidden state of the last time step, or the pooled result of all time step outputs as the comprehensive feature vector of the sentence;

[0085] When the sequence model is a Transformer model:

[0086] Use the self-attention mechanism to capture the long-distance dependencies between words based on the word vector order;

[0087] After the self-attention and feed-forward network of the encoder layer, the comprehensive feature vector of the sentence is output;

[0088] Furthermore, the extracting of context feature representation of text content data using a pre-trained language model comprises the following steps:

[0089] Step 21: input each sentence text into a pre-trained language model, encode the sentence text, and obtain a contextual semantic representation vector of each word in the sentence text; the pre-trained language model is a Bert model;

[0090] Step 22: Input the sequence of contextual semantic representation vectors of the word units of each sentence text into the Hierarchical Attention Network model to obtain a sentence-level context vector representation, which is the context feature representation;

[0091] Furthermore, the construction of a semantic relationship graph between sentences based on sentence feature vectors and context feature representations includes the following steps:

[0092] Step 31: Concatenate the sentence feature vector and context vector of each sentence text to form the sentence global feature vector of the sentence text; it can be understood that by fusing the sentence feature vector and the context vector, the semantic relationship graph can not only contain the contextual semantic information of the sentence, but also retain the word bag / word sequence information of the original text. This multimodal fusion representation can more comprehensively characterize the text semantics;

[0093] Step 32: For the sentence global feature vectors corresponding to any two sentence texts, calculate the cosine similarity of the two sentence global feature vectors. If the cosine similarity is greater than a preset similarity threshold, the two sentence texts are regarded as having a semantically related edge.

[0094] Step 33: Take each sentence as a sentence node, and connect two sentence nodes with semantically associated edges using a weighted edge; the weight of the weighted edge is the cosine similarity corresponding to the semantically associated edge;

[0095] Step 34: All sentence nodes and weighted edges between sentence nodes form a semantic relationship graph;

[0096] Furthermore, the collecting of the staffing domain knowledge base and the expansion of the semantic relationship graph based on the staffing domain knowledge include the following steps:

[0097] Step 41: Collect structured data from the field of staffing; the structured data includes but is not limited to job descriptions, skill requirements, salary information, etc.;

[0098] Step 42: extracting job triple relationships from structured data; importing the extracted triple relationships into a knowledge graph tool, constructing an ontology hierarchy and relationships, and obtaining a domain knowledge graph; the job triple relationship is, for example, (big data engineer, required skills, python), in which the backend development engineer and python respectively correspond to an entity node in the domain knowledge graph, and the required skills represent the connection relationship between the two entity nodes;

[0099] Step 43: Use a named entity recognition algorithm to perform named entity recognition on each sentence text in the text content data to obtain a target entity set;

[0100] Step 44: Find the connection path between the target entity sets of any two sentence texts in the domain knowledge graph, and based on the connection path, construct an extended weighted edge between every two sentence nodes in the semantic relationship graph, where the weight of the extended weighted edge is the inverse of the length of the connection path;

[0101] A simple example of expanding weighted edges is as follows:

[0102] Sentence 1: Zhang San is a software engineer with 5 years of Java development experience and is good at distributed

[0103] System design.

[0104] Sentence 2: Software engineer is one of the most in-demand positions at present, and having a solid programming foundation is crucial.

[0105] After performing entity recognition on sentences 1 and 2 respectively, the target entity sets obtained include "software engineer", "Java", and "distributed system";

[0106] In the domain knowledge graph, we may find a connection path: "Software Engineer - - Required Skills - -> Java", convert this relationship path into a weighted edge with a weight of 1, and add it to the semantic relationship graph to connect the sentence nodes corresponding to sentence 1 and sentence 2;

[0107] Furthermore, the feature fusion strategy is used to fuse the sentence feature vector and context feature representation of each sentence text through the expanded semantic relationship graph to generate the fused sentence feature vector in the following manner:

[0108] Inputting the expanded semantic relationship graph into a graph attention network, the graph attention network automatically learns the semantic relationship weight between every two sentence nodes;

[0109] For each sentence node, according to its relationship weight with other nodes and the sentence global feature vector of the sentence node, the graph attention network generates a new sentence vector representation that integrates context and knowledge. The new sentence vector representation is the fused sentence feature vector.

[0110] Specifically, the execution process of the graph attention network to generate a new sentence vector representation is as follows:

[0111] For each sentence node i, the graph attention network automatically learns its attention weight α_ij with other sentence nodes j; α_ij reflects how much sentence node j contributes to the semantics of sentence node i;

[0112] The Self-Attention mechanism of the graph attention network is used to adaptively calculate α_ij according to the node features. For sentence node i, the graph attention network generates a new sentence global feature vector h_i based on the sentence global feature vectors of all adjacent sentence nodes j and the corresponding attention weight α_ij. The new sentence global feature vector integrates the semantic information of the adjacent nodes of the node (for example, the new sentence global feature vector is calculated by weighted sum).

[0113] Furthermore, the method of clustering the fused sentence feature vectors to obtain K semantic clusters is as follows:

[0114] Use a clustering algorithm such as K-Means++ or a deep clustering network to perform cluster analysis on the fused sentence feature vectors of all sentence texts to obtain K semantic clustering clusters;

[0115] It can be understood that each semantic cluster includes a number of sentence texts, and the sentence texts in each semantic cluster have similar semantic characteristics;

[0116] Furthermore, the method of extracting the keywords representing each semantic cluster from each semantic cluster is as follows:

[0117] For each semantic cluster, calculate the average cosine similarity between its cluster centroid and the fused sentence feature vector of all the sentence texts in it;

[0118] The K1 semantic clusters with the highest average cosine similarity are taken as key clusters; K1 is a preset quantity parameter;

[0119] In each key cluster, the words are sorted according to the similarity score between the contextual semantic representation vector of each word and the centroid of the key cluster, and the first N1 words are used as the keywords of the key cluster;

[0120] Through the division of key clusters, the text content data is automatically classified, and keywords are extracted for each type of text data. For example, the semantics of semantic cluster 1 may be job responsibilities, and the keywords extracted from this semantic cluster may include: software development, project management, demand analysis, code review, etc. The semantics of semantic cluster 2 may be technical skills, and the keywords extracted from this semantic cluster may be Java, Python, distributed systems, databases, etc.

[0121] Example 2

[0122] like Figure 2 As shown, a content feature extraction system for a personnel configuration CMS includes a text data collection module, a semantic relationship graph construction module, a feature fusion module, and a keyword extraction module; wherein each module is electrically connected;

[0123] The text data collection module collects the text content data of the personnel configuring the CMS and sends the text content data to the semantic relationship graph construction module;

[0124] The semantic relationship graph construction module divides the text content data into sentences, obtains a sentence text set, performs word segmentation on each sentence text, converts each segmented word into a word vector, generates a sentence feature vector for each sentence text, extracts the context feature representation of the text content data using a pre-trained language model, constructs a semantic relationship graph between sentences based on the sentence feature vector and the context feature representation, collects a knowledge base in the field of staffing, expands the semantic relationship graph based on the knowledge in the field of staffing, and sends the semantic relationship graph to the feature fusion module;

[0125] The feature fusion module uses the feature fusion strategy to fuse the sentence feature vectors through the semantic relationship graph to generate a fused sentence feature vector, and sends the fused sentence feature vector of each sentence text to the keyword extraction module;

[0126] The keyword extraction module clusters the fused sentence feature vectors to obtain K semantic clusters, and extracts keywords representing each semantic cluster from each semantic cluster.

[0127] Example 3

[0128] Figure 3 Schematic diagram of a computer-readable storage medium structure provided by an embodiment of the present application. Figure 3As shown, it is a computer-readable storage medium 200 according to an embodiment of the present application. Computer-readable instructions are stored on the computer-readable storage medium 200. When the computer-readable instructions are executed by the processor, the personnel configuration CMS content feature extraction method according to the embodiment of the present application described with reference to the above figures can be executed. The computer-readable storage medium 200 includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0129] In addition, according to the implementation of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium, the non-transitory machine-readable storage medium stores machine-readable instructions, the machine-readable instructions can be run by a processor to execute instructions corresponding to the method steps provided by the present application, and when the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are executed.

[0130] The methods, apparatuses, and devices of the present application may be implemented in many ways. For example, the methods, apparatuses, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers recording media storing programs for executing the method according to the present application.

[0131] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0132] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0133] The above preset parameters or preset thresholds are all set by technicians in this field according to actual conditions or obtained through large-scale data simulation.

[0134] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for extracting content features of a personnel configuration CMS, characterized in that: The following steps are involved: Step 1: Collect text content data of personnel configuration CMS; Step 2: Segment the text content data into sentences to obtain a set of sentence texts, segment each sentence text, convert each segmented word into a word vector, and generate a sentence feature vector for each sentence text; Step 3: Use the pre-trained language model to extract the contextual feature representation of the text content data, and build a semantic relationship graph between sentences based on the sentence feature vector and the contextual feature representation; Step 4: Collect the knowledge base of staffing domain and expand the semantic relationship graph based on the knowledge of staffing domain; Step 5: Using the feature fusion strategy, the sentence feature vector and context feature representation of each sentence text are fused through the expanded semantic relationship graph to generate a fused sentence feature vector; Step 6: Cluster the fused sentence feature vector to obtain K semantic clusters; K is the number of semantic clusters obtained; Step 7: Extract keywords representing each semantic cluster from each semantic cluster; The method of extracting context feature representation of text content data using a pre-trained language model comprises the following steps: Step 21: input each sentence text into a pre-trained language model, encode the sentence text, and obtain a contextual semantic representation vector of each word in the sentence text; the pre-trained language model is a Bert model; Step 22: Input the sequence of contextual semantic representation vectors of the word units of each sentence text into the HierarchicalAttention Network model to obtain a contextual vector representation at the sentence level, which is the contextual feature representation; The method of constructing a semantic relationship graph between sentences based on sentence feature vectors and context feature representations includes the following steps: Step 31: concatenate the sentence feature vector and context vector of each sentence text to form a sentence global feature vector of the sentence text; Step 32: For the sentence global feature vectors corresponding to any two sentence texts, calculate the cosine similarity of the two sentence global feature vectors. If the cosine similarity is greater than a preset similarity threshold, the two sentence texts are regarded as having a semantically related edge. Step 33: Take each sentence as a sentence node, and connect two sentence nodes with semantically associated edges using a weighted edge; the weight of the weighted edge is the cosine similarity corresponding to the semantically associated edge; Step 34: All sentence nodes and weighted edges between sentence nodes form a semantic relationship graph.

2. A method for extracting content features of a personnel allocation CMS according to claim 1, characterized in that: The collecting of text content data of personnel configuring CMS comprises the following steps: Step 11: Identify the staffing CMS data sources and collect text data from the staffing CMS data sources; Step 12: Clean and label the text data to obtain text content data.

3. A method for extracting content features of a personnel allocation CMS according to claim 2, characterized in that: The method of segmenting the text content data into sentences and obtaining a sentence text set is as follows: Based on language rules, sentence boundaries of text content data are identified, and according to sentence boundaries, the text content data is divided into different sentence texts to form a sentence text set.

4. A method for extracting content features of a personnel allocation CMS according to claim 3, characterized in that: The method of segmenting each sentence text, converting each segmented word into a word vector, and generating a sentence feature vector for each sentence text is as follows: Use the word segmentation tool to divide each sentence text into words and obtain the word set corresponding to each sentence text; Use the pre-trained word vector model to convert each word in the word set into a word vector form; The word vectors of each word in the word set corresponding to each sentence text are sequentially combined into a word vector set; The sequence model is used to convert the word vector set into a comprehensive feature vector of a sentence, and the comprehensive feature vector of the sentence is the sentence feature vector.

5. A method for extracting content features of a personnel allocation CMS according to claim 4, characterized in that: The collecting of the staffing domain knowledge base and the expansion of the semantic relationship graph based on the staffing domain knowledge include the following steps: Step 41: Collect structured data from the field of staffing; the structured data includes but is not limited to job descriptions, skill requirements, salary information, etc.; Step 42: Extract job triple relationships from structured data; import the extracted triple relationships into a knowledge graph tool, construct an ontology hierarchy and relationship, and obtain a domain knowledge graph; Step 43: Use a named entity recognition algorithm to perform named entity recognition on each sentence text in the text content data to obtain a target entity set; Step 44: Find the connection path between the target entity sets of any two sentence texts in the domain knowledge graph, and based on the connection path, construct an extended weighted edge between every two sentence nodes in the semantic relationship graph, where the weight of the extended weighted edge is the inverse of the length of the connection path.

6. A method for extracting content features of a personnel allocation CMS according to claim 5, characterized in that: The feature fusion strategy is used to fuse the sentence feature vector and context feature representation of each sentence text through the expanded semantic relationship graph to generate the fused sentence feature vector in the following manner: Inputting the expanded semantic relationship graph into a graph attention network, the graph attention network automatically learns the semantic relationship weight between every two sentence nodes; For each sentence node, based on its relationship weight with other nodes and the sentence global feature vector of the sentence node, the graph attention network generates a new sentence vector representation that integrates context and knowledge. The new sentence vector representation is the fused sentence feature vector.

7. A method for extracting content features of a personnel allocation CMS according to claim 6, characterized in that: The method of extracting the keywords representing each semantic cluster from each semantic cluster is as follows: For each semantic cluster, calculate the average cosine similarity between its cluster centroid and the fused sentence feature vector of all the sentence texts in it; The K1 semantic clusters with the highest average cosine similarity are taken as key clusters; K1 is the preset quantity parameter; In each key cluster, the words are sorted according to the similarity score between the contextual semantic representation vector of each word and the centroid of the key cluster, and the first N1 words are used as the keywords of the key cluster.

8. A personnel configuration CMS content feature extraction system, which is used to implement the personnel configuration CMS content feature extraction method described in any one of claims 1 to 7, characterized in that: It includes a text data collection module, a semantic relationship graph construction module, a feature fusion module and a keyword extraction module; wherein each module is electrically connected; The text data collection module collects the text content data of the personnel configuring the CMS and sends the text content data to the semantic relationship graph construction module; The semantic relationship graph construction module divides the text content data into sentences, obtains a sentence text set, performs word segmentation on each sentence text, converts each segmented word into a word vector, generates a sentence feature vector for each sentence text, extracts the context feature representation of the text content data using a pre-trained language model, constructs a semantic relationship graph between sentences based on the sentence feature vector and the context feature representation, collects a knowledge base in the field of staffing, expands the semantic relationship graph based on the knowledge in the field of staffing, and sends the semantic relationship graph to the feature fusion module; The feature fusion module uses the feature fusion strategy to fuse the sentence feature vectors through the semantic relationship graph to generate a fused sentence feature vector, and sends the fused sentence feature vector of each sentence text to the keyword extraction module; The keyword extraction module clusters the fused sentence feature vectors to obtain K semantic clusters, and extracts keywords representing each semantic cluster from each semantic cluster.

9. A computer-readable storage medium, characterized in that: A rewritable computer program is stored thereon; When the computer program is executed on a computer device, the computer device executes the method for extracting content features of a personnel configuration CMS as described in any one of claims 1 to 7 in the background.

Images