Intelligent employment system based on artificial intelligence

Abstract

The invention belongs to the technical field of employment recommendation, and discloses an intelligent employment system based on artificial intelligence. Comprising the steps of collecting comprehensive data of job seekers; performing data preprocessing on the comprehensive data to generate a job seeker personal data set; based on the personal data set of the job seeker, extracting keywords of the job seeker by using a natural language processing method; calculating professional ability scores of the job seeker in n fields according to the keywords; based on the professional ability scores of the job seekers in the n fields, generating a job seeker occupational recommendation list by using a Bayesian network algorithm; and the job seeker occupation recommendation list is subjected to secondary screening to obtain a final intelligent employment scheme, so that the employment matching degree and satisfaction degree of the job seekers are greatly improved, and the supply and demand contradiction of the current employment market can be relieved.

Description

Technical Field

[0001] The present invention relates to the field of employment recommendation technology, and more specifically, to an intelligent employment system based on artificial intelligence. Background Art

[0002] The patent with application publication number CN116402485A discloses a smart city employment management method and an Internet of Things system, medium and device; including a user platform, a service platform, and a management platform; the smart city employment management method is executed by the management platform of the smart city employment management Internet of Things system, including: obtaining company-related information; determining the target industry and target company based on the company-related information and combined with the query conditions; predicting the target company's turnover rate and entry rate based on the target company's company-related information; returning the target company's turnover rate and entry rate to the user platform through the service platform; and generating a first warning message based on the turnover rate that meets the first preset requirement, and / or generating a second warning message based on the entry rate that meets the second preset requirement, and returning the first warning message and / or the second warning message to the user platform based on the service platform.

[0003] However, in the current job market, there is still a serious information asymmetry problem between job seekers and employers; traditional job search channels and recommendation systems have many shortcomings, making it difficult to effectively match job seekers and positions, resulting in inefficient allocation of employment resources; first, the traditional method does not consider enough data in collecting and processing job seeker information, ignores the comprehensive data of job seekers, such as skills, specialties, personal hobbies, etc., and cannot fully grasp the professional ability and career inclination of job seekers; in addition, it can usually only process structured data and cannot fundamentally extract valuable information, resulting in information loss; secondly, the traditional method is difficult to quantify the professional ability level of job seekers in different fields in analyzing job seeker data; this extensive data processing method greatly reduces the analysis quality and value mining ability of job seeker data; thirdly, the traditional method also has shortcomings in job seeker career recommendations; they usually adopt simple rule matching or recommendation models based on limited features, which cannot fully consider the matching degree between the multi-dimensional characteristics of job seekers and positions, resulting in inaccurate recommendation results and inability to truly meet the employment needs of job seekers, which in turn leads to aggravated information asymmetry between job seekers and employers and inefficient allocation of employment resources.

[0004] In view of this, the present invention proposes an intelligent employment system based on artificial intelligence to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions: an intelligent employment system based on artificial intelligence, comprising: a data collection and processing module, for collecting comprehensive data of job seekers; and pre-processing the comprehensive data to generate a personal data set of job seekers;

[0006] The professional competence assessment module uses natural language processing methods to extract job seekers’ keywords based on their personal data sets and calculates the job seekers’ professional competence scores in n areas based on the keywords.

[0007] The preliminary recommendation module generates a list of recommended careers for job seekers based on their professional ability scores in n fields using the Bayesian network algorithm;

[0008] The career optimization module is used to conduct a secondary screening of the job seeker's career recommendation list to obtain the final smart employment plan; each module is connected by wired and / or wireless means.

[0009] Furthermore, the comprehensive data includes basic personal information, work experience, skills and expertise, award certificates and personal hobbies.

[0010] Furthermore, the method of preprocessing the comprehensive data includes:

[0011] Remove noise data from the comprehensive data, the noise data includes irrelevant characters and blank characters; obtain preliminary unstructured data; convert the preliminary unstructured data into structured data, the structured data includes categorical variables and text data; perform word embedding encoding on the categorical variables; obtain categorical features; perform feature extraction on the text data to obtain text features; integrate the categorical features and text features to construct a unified data set, which is the job seeker's personal data set.

[0012] Furthermore, the method of converting the preliminary unstructured data into structured data includes:

[0013] Construct a hidden Markov model and use it to identify entities from unstructured data and label the entities as corresponding categorical variables. The entities include names of people, places, organizations and time.

[0014] The method of identifying entities from unstructured data includes:

[0015] Perform feature extraction preprocessing on the preliminary unstructured data, input the preliminary unstructured data after feature extraction preprocessing into the constructed hidden Markov model, and decode it to obtain the final hidden state sequence, which is the final identified entity;

[0016] Convert the preliminary unstructured data into a plain text format to obtain plain text data; count all unique terms appearing in the plain text data, map each unique term to an index number, and construct a dictionary; convert the dictionary into a bag-of-words vector representation, where each element of the bag-of-words vector records the number of times the corresponding unique term appears;

[0017] Define the number of potential topics R; use the bag-of-words vector as input, and based on the topic model, fit the model parameters of the topic model through gibbs sampling or variational inference algorithm. The model parameters include the word distribution of each potential topic and the topic distribution of each preliminary unstructured data;

[0018] For each potential topic, check the N terms with the largest probability values ​​in its word distribution and use them as the focus words of the potential topic; assign a label to each potential topic;

[0019] For the preliminary unstructured data, check the B potential topics with the largest probability values ​​in the topic distribution; these potential topics are the main topics to which the preliminary unstructured data belongs; combine the focus words of each main topic into a piece of text as the text representation of the main topic; the text representations of all main topics constitute the text data.

[0020] Furthermore, the method of constructing the hidden Markov model includes:

[0021] Collect m groups of text data with entity annotations as training data; extract training features from the training data; the training features include the current word, part of speech, case mode and whether it contains numbers; use the training features as observations of the hidden Markov model; encode the training features;

[0022] Define a tag sequence based on training data to represent entity types; use the tag sequence as the hidden state sequence of the hidden Markov model;

[0023] Initialize the parameters of the hidden Markov model, including the initial state probability, state transition probability and emission probability;

[0024] The initial state probability π(1) represents the probability of the hidden state at time t=1; the state transition probability A(i,j) represents the probability of transitioning from hidden state i to hidden state j at time t; the emission probability B(j,k) represents the probability of observing the observation value o_k when in hidden state j; where k is the index of the observation value;

[0025] Using the initialized parameters, the hidden Markov model is trained on the training data; and the Baum-Welch algorithm is used to maximize the log-likelihood of the hidden Markov model on the training data; the optimal parameters of the hidden Markov model are obtained, and the parameters at this time are fixed; and the constructed hidden Markov model is obtained.

[0026] Furthermore, the method of decoding to obtain the final hidden state sequence includes:

[0027] The preliminary unstructured data is constructed into a preliminary observation sequence O of length T, and the preliminary observation sequence O is divided into M segments, each of which is

[0028] Define Q(l,k) to represent the kth optimal path of the lth segment, and P(l,k) to represent the probability corresponding to Q(m,k); for the first segment, use the Viterbi algorithm to solve the K optimal paths Q(1,k) and the corresponding probabilities P(1,k) of the first L observations, where 1≤k≤K; K refers to the number of optimal paths that need to be retained in each segment;

[0029] For each subsequent segment l, use the Viterbi algorithm to solve the K optimal paths Q(l,k) and corresponding probabilities P(l,k) of the segment observations;

[0030] For each optimal path Q(l,k), calculate its combined probability with all K optimal paths in the previous segment; the calculation formula for the combined probability is:

[0031] C(J,k)=P(l-1,J)×P(l,k)×∑α(t)×P′(q(l-1,J)_L→q(l,k)_1|f(l-1,J),f(l,k),o_L); where C(J,k) represents the probability of the Jth optimal path of the l-1th segment and the kth path of the lth segment to form a complete path, P(l-1,J) represents the probability of the Jth optimal path of the l-1th segment, P(l,k) represents the probability of the kth optimal path of the lth segment, and α( t) is the attention weight at time t; f(l,k) is the feature vector of the kth optimal path of the lth segment, f(l-1,J) is the feature vector of the Jth optimal path of the l-1th segment, and o_L is the observation value at the junction of the two segments; q(l-1,J)_L→q(l,k)_1 means the transition from the last state q(l-1,J)_L of the Jth optimal path of the l-1th segment to the first state q(m,k)_1 of the kth optimal path of the lth segment; P′() is the state transition probability;

[0032] Keep the K combinations of optimal paths with the largest combined probability C(j,k). Among the K optimal paths of the last segment M, select the optimal path Q(M,k) with the largest probability P(M,k) as the final hidden state sequence. Q(M,k) is the kth optimal path representing the Mth segment, and P(M,k) is the probability corresponding to Q(M,k).

[0033] Furthermore, the keyword extraction method includes:

[0034] Convert the job seeker personal data set into text format to obtain a text set; define a window with a size of r, slide the window in the text set, record the co-occurrence relationship of words in the window, represent each unique word as a node, and if two words co-occur in the window, connect them with an edge, and the weight of the edge is the number of their co-occurrences, to obtain a weighted directed graph, recorded as a word co-occurrence graph;

[0035] In the word co-occurrence graph, calculate the importance score of each node; the word co-occurrence graph G = (V, E), where V is the set of nodes and E is the set of edges;

[0036] Define the adjacency matrix A. If there is an edge from node a to node b, then the element A[a,b] in the adjacency matrix is ​​1, otherwise it is 0;

[0037] For each node a, calculate its comprehensive out-degree d(a) = ∑ b w(a,b)×A[a,b]×pos(a,b); where w(a,b) is the weight of the edge from node a to node b, and pos(a,b) is the part-of-speech weight function of nodes a and b;

[0038] pos(a,b)=match(ps(a),ps(b))×γ+(1-match(ps(a),ps(b)))×ε; where ps(a) is the part-of-speech tag of node a, and ps(b) is the part-of-speech tag of node b; γ and ε are two weight parameters; match(ps(a),ps(b)) is the matching function, which is 1 when ps(a) and ps(b) are the same, otherwise it is 0;

[0039] Initialize the importance score W(a) of each node a as Where D is the total number of nodes. In each iteration, the importance score W(a) of each node a is updated according to the importance scores of other nodes.

[0040] Among them, S(e) is the value of the damping coefficient at the e-th iteration, W(b) is the importance score of node b, and d(b) is the comprehensive out-degree of node b;

[0041] Where S(e-1) is the value of the damping coefficient at the e-1th iteration, θ is a positive learning rate that controls the amplitude of the adjustment, W(a)_(e) is the importance score of node a at the e-th iteration, and W(a)_(e-1) is the importance score of node a at the e-1th iteration;

[0042] Repeat the iteration until the importance score of each node converges, that is, the score change between two consecutive iterations is less than the preset change threshold; output the importance score of each node at the last iteration, sort the nodes from high to low according to the importance score, and select the words corresponding to the first H nodes as keywords.

[0043] Furthermore, the calculation method of the professional ability score includes:

[0044] Define a domain set, including n1 domains, each representing a domain; determine a set of keywords X_c = {x_I} for each domain c, where x_I is the Ith keyword of domain c;

[0045] Construct an n1×m1 domain-keyword matrix ME, where n1 is the number of domains, m1 is the total number of keywords in all domains, and each element ME_cI of the domain-keyword matrix ME represents the importance weight of the I-th keyword in domain c; construct a job seeker keyword vector Y based on the keywords in the job seeker's personal data set, where the element Y_y=1 in Y indicates that the set of keywords in the job seeker's personal data set contains the y-th keyword, otherwise Y_y=0;

[0046] Multiply the job applicant keyword vector Y with the field-keyword matrix ME to obtain the job applicant-field vector SH, where the element SH_c in SH represents the job applicant's professional ability score in field c.

[0047] Furthermore, the method of generating the job seeker's occupation recommendation list includes:

[0048] Construct a Bayesian network model, define network nodes in the Bayesian network model, the network nodes include nodes mapped to the professional ability scores of job seekers in various fields and nodes mapped to occupations; nodes mapped to the professional ability scores of job seekers in various fields are recorded as professional ability score nodes, and nodes mapped to occupations are recorded as occupation recommendation nodes; conditional probability distributions between given network nodes are given;

[0049] Normalize the job seeker's professional ability scores in n fields, encode the occupations, and map them into discrete state values;

[0050] For each career recommendation node h, calculate its conditional probability Z(h|S1,S2,...,Sn); where S1,S2,...,Sn are the first professional ability score node to the nth professional ability score node, that is, all professional ability score nodes;

[0051] Where Z(S1) is the conditional probability distribution of S1, Z(S2) is the conditional probability distribution of S2, and Z(Sn) is the conditional probability distribution of Sn; Z(h,S1,S2,...,Sn) represents the joint probability distribution of the career recommendation node h and all professional ability score nodes S1,S2,...,Sn;

[0052] For each occupation, the corresponding conditional probability is calculated, and the occupations are sorted in descending order according to the size of the conditional probability. The first N2 occupations are selected to form a recommended occupation list for job seekers.

[0053] Furthermore, the secondary screening method includes:

[0054] Construct a membrane computing model and define the membrane structure of the membrane computing model, including the membrane hierarchy, membrane initialization objects and membrane computing rules; the membrane hierarchy includes the outermost layer, the middle layer and the innermost layer, the outermost layer is the skin membrane, the middle layer is the working membrane, and the innermost layer is the job seeker membrane;

[0055] Initialize the skin membrane to an empty set, the job membrane to a job seeker's occupation recommendation list, and the job seeker membrane to a job seeker's personal dataset;

[0056] The membrane calculation rules include sending rules, matching rules, screening rules and optimization rules; the sending rules are to send the keywords and professional ability scores of job seekers from the job seeker membrane to the job membrane; the matching rules are to calculate the comprehensive matching score of each occupation in the job seeker's occupation recommendation list in the job membrane according to the keywords and professional ability scores of the job seeker;

[0057] The match score is calculated by:

[0058] For each occupation v in the job seeker's occupation recommendation list, construct a occupation description vector V_v={u_ov}, where u_ov represents the importance value of the oth keyword to occupation v;

[0059] For job seekers, all their keywords are concatenated at the vector level to obtain the job seeker's comprehensive key vector Y′;

[0060] For occupation v, calculate the similarity between the comprehensive key vector of job seeker w and the occupation description vector as the keyword matching score Among them, ||Y′|| and ||V_v|| represent the L2 norm of Y′ and V_v respectively;

[0061] For occupation v, define the ability weights of the job seeker's professional ability scores in n fields respectively; then calculate the ability matching score between job seeker w and occupation v;

[0062] The keyword matching score and the ability matching score are weighted and summed to obtain the comprehensive matching score between job seeker w and occupation v;

[0063] The screening rule is to send occupations with comprehensive matching scores greater than or equal to the preset screening threshold from the work membrane to the skin membrane; the optimization rule is to further screen the occupations in the skin membrane;

[0064] The membrane calculation process is executed according to the defined membrane structure and membrane calculation rules. The membrane calculation process is an iterative process, and the membrane calculation rules are executed once at each time step. The membrane calculation process is continuously executed until the termination condition is met and the membrane calculation process is terminated. When the membrane calculation process is terminated, the object in the skin membrane is the final smart employment plan. The smart employment plan includes recommended occupations and comprehensive matching scores.

[0065] The technical effects and advantages of the intelligent employment system based on artificial intelligence of the present invention are as follows:

[0066] The present invention collects comprehensive data of job seekers comprehensively, and uses natural language processing and topic models to extract valuable information from unstructured data, so as to maximize the potential characteristics and professional capabilities of job seekers, and lay a solid data foundation for subsequent analysis; secondly, a keyword extraction algorithm based on a graph model and a professional ability evaluation method based on a matrix are adopted, which can accurately identify the core keywords of job seekers and quantify their professional ability levels in different fields, and provide an important basis for accurate matching of occupations; furthermore, a Bayesian network model is introduced to comprehensively consider the multi-dimensional characteristics of job seekers, such as keywords, professional ability scores, etc., and the matching degree between different occupations, so as to generate a high-quality occupation recommendation list, which greatly improves the accuracy and matching degree of the recommendation; finally, an optimization algorithm is used to optimize the preliminary occupation recommendation list, and by setting reasonable rules and optimization strategies, the quality and interpretability of the recommendation results can be further improved, and a more suitable smart employment plan can be provided for job seekers; the employment matching degree and satisfaction of job seekers are greatly improved, which helps to alleviate the contradiction between supply and demand in the current employment market. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A schematic diagram of an intelligent employment system based on artificial intelligence of the present invention;

[0068] Figure 2 This is a schematic diagram of the artificial intelligence-based smart employment method of the present invention. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not 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.

[0070] Example 1

[0071] See also Figure 1 As shown, the smart employment system based on artificial intelligence described in this embodiment includes:

[0072] The data collection and processing module is used to collect the comprehensive data of job seekers; and pre-process the comprehensive data to generate a personal data set of job seekers;

[0073] The professional competence assessment module uses natural language processing methods to extract job seekers’ keywords based on their personal data sets and calculates the job seekers’ professional competence scores in n areas based on the keywords.

[0074] The preliminary recommendation module generates a list of recommended careers for job seekers based on their professional ability scores in n fields using the Bayesian network algorithm;

[0075] The career optimization module is used to conduct a secondary screening of the job seeker's career recommendation list to obtain the final intelligent employment plan; each module is connected by wired and / or wireless means to achieve data transmission between modules.

[0076] Furthermore, the comprehensive data includes basic personal information (name, age, gender, educational background), work experience, skills and expertise, award certificates and personal hobbies, etc.

[0077] The comprehensive data is collected through various channels, including filling out forms online, on-site interviews, and resume uploads, etc.; and the collected comprehensive data is unstructured (free text).

[0078] Furthermore, the method of preprocessing the comprehensive data includes:

[0079] Noise data is removed from the comprehensive data, the noise data including irrelevant characters and blank characters; and preliminary unstructured data is obtained; the irrelevant characters include special characters that cannot be recognized by a computer, such as some unprintable characters or garbled characters.

[0080] Convert preliminary unstructured data into structured data, including categorical variables and text data; perform word embedding encoding on categorical variables (such as education, gender, etc.); obtain categorical features; perform feature extraction (such as TF-IDF) on text data (such as work experience) to obtain text features; integrate categorical features and text features to construct a unified data set, which is the job seeker’s personal data set.

[0081] The method of converting the preliminary unstructured data into structured data includes:

[0082] A hidden Markov model is constructed and used to identify entities from unstructured data. The entities are labeled as corresponding categorical variables. The entities include names of people, places, organizations and time.

[0083] Ways to construct a hidden Markov model include:

[0084] Collect m groups of text data with entity annotations as training data; extract training features from the training data; the training features include the current word, part of speech, case mode and whether it contains numbers; use the training features as observations of the hidden Markov model; encode the training features, such as one-hot encoding.

[0085] A tag sequence is defined based on the training data to indicate the entity type, such as B-PER (beginning of a person's name), I-PER (inside a person's name), B-ORG (beginning of an organization), etc.; the tag sequence is used as a hidden state sequence (a sequence consisting of hidden states) of a hidden Markov model; for example, defining a tag sequence means defining a label sequence for entities in the training data (such as names of people, places, names of organizations, etc.) to indicate the type of the entity. Specifically, the BIO labeling method is used, wherein B-XXX indicates the beginning of entity XXX; I-XXX indicates the middle part of entity XXX; and O indicates that it does not belong to any entity.

[0086] For example, for a sentence "I am studying at Peking University", the following tag sequence is defined:

[0087] I / O am studying in / OBeijing / B-LOC University / I-LOC / O; B-LOC indicates that "Beijing" is the beginning of a place name entity, and I-LOC indicates that "University" is the middle part of the place name entity.

[0088] Initialize the parameters of the hidden Markov model, including the initial state probability, state transition probability and emission probability.

[0089] The initial state probability π(1) represents the probability of the hidden state at time t=1; the state transition probability A(i,j) represents the probability of transitioning from hidden state i to hidden state j at time t; the emission probability B(j,k) represents the probability of observing the observation value o_k in the hidden state j; where k is the index of the observation value.

[0090] The state transition probability is estimated using the frequencies of all pairs of adjacent hidden states in the training data, and the emission probability is estimated using the frequency of observations under each hidden state in the training data.

[0091] Using the initialized parameters, the hidden Markov model is trained on the training data; and the Baum-Welch algorithm is used to maximize the log-likelihood of the hidden Markov model on the training data; the optimal parameters of the hidden Markov model are obtained, and the parameters at this time are fixed; and the constructed hidden Markov model is obtained.

[0092] The preliminary unstructured data is preprocessed with feature extraction (in the same way as extracting training features), and the preliminary unstructured data after feature extraction preprocessing is input into the constructed hidden Markov model, and decoded to obtain the final hidden state sequence, which is the final identified entity.

[0093] Specifically, the preliminary unstructured data is constructed into a preliminary observation sequence O of length T, and the preliminary observation sequence O is divided into M segments, each segment length (number of observations) is

[0094] Define Q(l,k) to represent the kth optimal path of the lth segment, and P(l,k) to represent the probability corresponding to Q(m,k); for the first segment (l=1), use the Viterbi algorithm to solve the K optimal paths Q(1,k) and the corresponding probabilities P(1,k) of the first L observations, where 1≤k≤K; K refers to the number of optimal paths (most likely hidden state sequences) that need to be retained in each segment.

[0095] For each subsequent segment l, (2≤l≤M), the Viterbi algorithm is used to solve the K optimal paths Q(l,k) and corresponding probabilities P(l,k) of the observed values ​​of the segment.

[0096] For each optimal path Q(l,k), calculate its combined probability with all K optimal paths in the previous segment; the calculation formula for the combined probability is:

[0097] C(J,k)=P(l-1,J)×P(l,k)×∑α(t)×P′(q(l-1,J)_L→q(l,k)_1|f(l-1,J),f(l,k),o_L); where C(J,k) represents the probability of the Jth optimal path of the l-1th segment and the kth path of the lth segment forming a complete path, P(l-1,J) represents the probability of the Jth optimal path of the l-1th segment (it is the probability value of the path calculated by the Viterbi algorithm or other methods under the observed value of the l-1th segment), P(l,k) represents the probability of the kth optimal path of the lth segment (it is the probability value of the path calculated by the Viterbi algorithm or other methods under the observed value of the lth segment), and α(t) is the probability of the kth optimal path of the lth segment at time t. The attention weights are obtained through adaptive learning of the neural network; f(l,k) is the feature vector of the kth optimal path of the lth segment (some features are extracted from the optimal path to form a feature vector, the features include the length of the path, the number of times the entity type appears, and the word frequency of the observation value), f(l-1,J) is the feature vector of the Jth optimal path of the l-1th segment, o_L is the observation value at the junction of the two segments, that is, the last observation value of the l-1th segment and the first observation value of the lth segment; q(l-1,J)_L→q(l,k)_1 means the transition from the last state q(l-1,J)_L of the Jth optimal path of the l-1th segment to the first state q(m,k)_1 of the kth optimal path of the lth segment; P′() is the state transition probability.

[0098] Keep the K combinations of optimal paths with the largest combined probability C(j,k). Among the K optimal paths of the last segment M, select the optimal path Q(M,k) with the largest probability P(M,k) as the final hidden state sequence. Q(M,k) is the kth optimal path representing the Mth segment, and P(M,k) is the probability corresponding to Q(M,k).

[0099] Topic models are used to discover potential topics from unstructured data, and all potential topics constitute text data. Specifically, the preliminary unstructured data is converted into a plain text format to obtain plain text data. All unique terms appearing in the plain text data are counted, each unique term is mapped to an index number, and a dictionary is constructed. The dictionary is converted into a bag-of-words vector representation, and each element of the bag-of-words vector records the number of times the corresponding unique term appears.

[0100] Define the number of potential topics R; use heuristics or cross-validation to determine a reasonable R value;

[0101] Using bag-of-words vectors as input, based on the topic model, the model parameters of the topic model are fitted by gibbs sampling or variational inference algorithm, and the model parameters include the word distribution of each potential topic and the topic distribution of each preliminary unstructured data; the backbone algorithm of the topic model is LDA or Latent Dirichlet Allocation model.

[0102] For each potential topic, check the N terms with the largest probability values ​​in its word distribution and use them as the focus words of the potential topic; assign a label to each potential topic; manually analyze and interpret the semantic meaning of each topic.

[0103] For the preliminary unstructured data, check the B potential topics with the largest probability values ​​in the topic distribution; these potential topics are the main topics to which the preliminary unstructured data belongs; combine the focus words of each main topic into a piece of text as the text representation of the main topic; the text representations of all main topics constitute the text data.

[0104] Furthermore, the keyword extraction methods include:

[0105] Convert the job applicant personal data set into text format to obtain a text set; define a window with a size of r (such as 5 words), slide the window in the text set, record the co-occurrence relationship of the words in the window, represent each unique word as a node, and if two words co-occur in the window, connect them with an edge, and the weight of the edge is the number of their co-occurrences, to obtain a weighted directed graph, recorded as a word co-occurrence graph.

[0106] In the word co-occurrence graph, the importance score of each node (word) is calculated; specifically, the word co-occurrence graph G = (V, E), where V is a set of nodes (words) and E is a set of edges.

[0107] Define the adjacency matrix A. If there is an edge from node a to node b, the element A[a,b] in the adjacency matrix is ​​1, otherwise it is 0.

[0108] For each node a, calculate its comprehensive out-degree d(a) = ∑ b w(a,b)×A[a,b]×pos(a,b); where w(a,b) is the weight of the edge from node a to node b, and pos(a,b) is the part-of-speech weight function of nodes a and b.

[0109] pos(a,b)=match(ps(a),ps(b))×γ+(1-match(ps(a),ps(b)))×ε; where ps(a) is the part-of-speech tag of node a, and ps(b) is the part-of-speech tag of node b, such as noun, verb, adjective, etc.; γ and ε are two weight parameters that control the weights of edges with the same part of speech and different parts of speech, and γ>ε is usually set; match(ps(a),ps(b)) is the matching function, which is 1 when ps(a) and ps(b) are the same, otherwise it is 0.

[0110] Initialize the importance score W(a) of each node a as Where D is the total number of nodes. In each iteration, the importance score W(a) of each node a is updated according to the importance scores of other nodes.

[0111] Where S(e) is the value of the damping coefficient at the e-th iteration, W(b) is the importance score of node b, and d(b) is the comprehensive out-degree of node b (calculated in the same way as d(a)).

[0112] Where S(e-1) is the damping coefficient (a value between 0 and 1, used to control the update rate of the node importance score) at the e-1th iteration, θ is a positive learning rate that controls the amplitude of the adjustment, W(a)_(e) is the importance score of node a at the e-th iteration, and W(a)_(e-1) is the importance score of node a at the e-1th iteration.

[0113] The iterations are repeated until the importance score of each node converges, that is, the score change between two consecutive iterations is less than a preset change threshold (set to a small value, such as 1e-5 or 1e-6).

[0114] Output the importance score of each node at the last iteration, sort the nodes from high to low according to the importance score, and select the words corresponding to the first H nodes as keywords.

[0115] Furthermore, the calculation method of professional ability score includes:

[0116] Define a domain set, including n1 domains, each representing a field, such as software development, marketing, financial accounting, etc.; determine a set of keywords X_c = {x_I} for each domain c, where x_I is the Ith keyword of domain c.

[0117] Construct an n1×m1 domain-keyword matrix ME, where n1 is the number of domains, m1 is the total number of keywords in all domains, and each element ME_cI of the domain-keyword matrix ME represents the importance weight of the Ith keyword in domain c.

[0118] According to the keywords in the job seeker's personal data set, a job seeker keyword vector Y is constructed, where an element Y_y=1 in Y indicates that the set of keywords in the job seeker's personal data set contains the yth keyword, otherwise Y_y=0.

[0119] Multiply the job applicant keyword vector Y with the field-keyword matrix ME to obtain the job applicant-field vector SH, where the element SH_c in SH represents the job applicant's professional ability score in field c.

[0120] It should be noted that the weights in the domain-keyword matrix ME are automatically learned through machine learning to more accurately reflect the importance of each keyword in the corresponding domain.

[0121] Furthermore, the method of generating the job seeker's occupation recommendation list includes:

[0122] A Bayesian network model is constructed and network nodes in the Bayesian network model are defined. The network nodes include nodes mapping the professional ability scores of job seekers in various fields and nodes mapping occupations; the nodes mapping the professional ability scores of job seekers in various fields are recorded as professional ability score nodes, and the nodes mapping occupations are recorded as occupation recommendation nodes.

[0123] Given a conditional probability distribution between network nodes (with enough data samples, use a machine learning algorithm to estimate the conditional probability distribution from the data).

[0124] The job applicants' professional ability scores in n fields are normalized, and the occupations are encoded and mapped into discrete state values.

[0125] For each career recommendation node h, calculate its conditional probability Z(h|S1,S2,...,Sn); where S1,S2,...,Sn are the first professional ability score node to the nth professional ability score node, that is, all professional ability score nodes.

[0126] Where Z(S1) is the conditional probability distribution of S1, Z(S2) is the conditional probability distribution of S2, and Z(Sn) is the conditional probability distribution of Sn; Z(h,S1,S2,...,Sn) represents the joint probability distribution of the career recommendation node h and all professional ability score nodes S1,S2,...,Sn (describing the probability that multiple random variables take certain values ​​at the same time).

[0127] For each occupation, the corresponding conditional probability is calculated, and the occupations are sorted in descending order according to the size of the conditional probability. The first N2 occupations are selected to form a recommended occupation list for job seekers.

[0128] Furthermore, the secondary screening method includes:

[0129] Construct a membrane computing model and define the membrane structure of the membrane computing model including the membrane hierarchy, membrane initialization objects and membrane computing rules; the membrane hierarchy includes the outermost layer, the middle layer and the innermost layer, the outermost layer is the skin membrane, the middle layer is the working membrane, and the innermost layer is the job seeker membrane.

[0130] Initialize the skin membrane to an empty set, the job membrane to a job seeker's occupation recommendation list, and the job seeker membrane to a job seeker's personal dataset.

[0131] The membrane calculation rules include sending rules, matching rules, screening rules and optimization rules; the sending rules are to send the job seeker's keywords and professional ability scores from the job seeker membrane to the job membrane; the matching rules are to calculate the comprehensive matching score of each occupation in the job seeker's occupation recommendation list in the job membrane according to the job seeker's keywords and professional ability scores.

[0132] The match score is calculated by:

[0133] For each occupation v in the job seeker's occupation recommendation list, construct a occupation description vector V_v={u_ov}, where u_ov represents the importance value of the oth keyword to occupation v;

[0134] For job seekers, all their keywords are concatenated at vector level to obtain the job seeker's comprehensive key vector Y′.

[0135] For occupation v, calculate the similarity between the comprehensive key vector of job seeker w and the occupation description vector as the keyword matching score Among them, ||Y′|| and ||V_v|| represent the L2 norm of Y′ and V_v respectively.

[0136] For occupation v, the ability weights of the job seeker's professional ability scores in n fields are defined respectively, which represent the importance of the occupation to the ability in each field, and are obtained by analyzing a large amount of job position data; then the ability matching score between job seeker w and occupation v is calculated (that is, the ability weight multiplied by the professional ability score).

[0137] The keyword matching score and the ability matching score are weighted and summed to obtain the comprehensive matching score between job seeker w and occupation v; the weights of the weighted summation control the importance of keywords and abilities in the matching score, and the optimal value is determined by the cross-validation method.

[0138] The screening rule is to send occupations with comprehensive matching scores greater than or equal to the preset screening threshold from the work membrane to the skin membrane; the optimization rule is to further screen the occupations in the skin membrane (iteratively execute the membrane calculation process) to obtain the final smart employment plan; preset an initial screening threshold based on experience, such as 0.7; within a certain data range, perform a grid search on the screening threshold, evaluate the result quality under different thresholds, and select the optimal screening threshold.

[0139] The membrane calculation process is executed according to the defined membrane structure and membrane calculation rules. The membrane calculation process is an iterative process, and the membrane calculation rules are executed once at each time step. The membrane calculation process is executed continuously until the termination condition is met (such as reaching the maximum number of iterations or the number of objects in the skin membrane is stable), and the membrane calculation process is terminated. When the membrane calculation process terminates, the object (occupation) in the skin membrane is the final smart employment plan. The smart employment plan includes recommended occupations and comprehensive matching scores.

[0140] In this embodiment, by comprehensively collecting comprehensive data of job seekers and extracting valuable information from unstructured data using natural language processing and topic models, the potential characteristics and professional capabilities of job seekers are explored to the maximum extent, laying a solid data foundation for subsequent analysis; secondly, a keyword extraction algorithm based on a graph model and a professional ability assessment method based on a matrix are adopted, which can accurately identify the core keywords of job seekers and quantify their professional ability levels in different fields, providing an important basis for accurate matching of occupations; furthermore, a Bayesian network model is introduced to comprehensively consider the multi-dimensional characteristics of job seekers, such as keywords, professional ability scores, etc., and the matching degree between different occupations, so as to generate a high-quality occupation recommendation list, which greatly improves the accuracy and matching degree of the recommendation; finally, an optimization algorithm is used to optimize the preliminary occupation recommendation list, and by setting reasonable rules and optimization strategies, the quality and interpretability of the recommendation results can be further improved, providing job seekers with a more suitable smart employment plan; the employment matching degree and satisfaction of job seekers are greatly improved, which helps to alleviate the contradiction between supply and demand in the current employment market.

[0141] Example 2

[0142] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides an intelligent employment method based on artificial intelligence, including:

[0143] S1. Collect comprehensive data of job seekers; and pre-process the comprehensive data to generate a personal data set of job seekers;

[0144] S2. Based on the job seeker’s personal data set, the job seeker’s keywords are extracted using natural language processing methods; and based on the keywords, the job seeker’s professional ability scores in n fields are calculated;

[0145] S3, based on the job seeker's professional ability scores in n fields, using the Bayesian network algorithm to generate a list of recommended careers for the job seeker;

[0146] S4. Conduct a secondary screening of the job seeker's career recommendation list to obtain the final smart employment plan.

[0147] Example 3

[0148] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the above-mentioned artificial intelligence-based smart employment method is implemented.

[0149] Since the electronic device introduced in this embodiment is an electronic device used to implement the smart employment method based on artificial intelligence in the embodiment of this application, based on the smart employment method based on artificial intelligence introduced in the embodiment of this application, the technical personnel of this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application is not introduced in detail here. As long as the technical personnel of this field implement the electronic device used in the smart employment method based on artificial intelligence in the embodiment of this application, it belongs to the scope of protection of this application.

[0150] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0151] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. The intelligent employment system based on artificial intelligence is characterized by: include: Data collection and processing module, used to collect comprehensive data of job seekers; The comprehensive data is preprocessed to generate a personal data set of job seekers; The professional competence assessment module uses natural language processing methods to extract job seekers’ keywords based on their personal data sets; And based on the keywords, calculate the job seeker's professional ability scores in n fields; The preliminary recommendation module generates a list of recommended careers for job seekers based on their professional ability scores in n fields using the Bayesian network algorithm; The career optimization module is used to conduct a secondary screening of the job seeker's career recommendation list to obtain the final smart employment plan; each module is connected by wired and / or wireless means.

2. The intelligent employment system based on artificial intelligence according to claim 1 is characterized in that: The comprehensive data includes basic personal information, work experience, skills and expertise, award certificates and personal hobbies.

3. The intelligent employment system based on artificial intelligence according to claim 2 is characterized in that: The method of preprocessing the comprehensive data includes: Remove noise data from the comprehensive data, the noise data includes irrelevant characters and blank characters; obtain preliminary unstructured data; convert the preliminary unstructured data into structured data, the structured data includes categorical variables and text data; perform word embedding encoding on the categorical variables; obtain categorical features; perform feature extraction on the text data to obtain text features; integrate the categorical features and text features to construct a unified data set, which is the job seeker's personal data set.

4. The intelligent employment system based on artificial intelligence according to claim 3 is characterized in that: The method of converting the preliminary unstructured data into structured data includes: Construct a hidden Markov model and use it to identify entities from unstructured data and label the entities as corresponding categorical variables. The entities include names of people, places, organizations and time. The method of identifying entities from unstructured data includes: Perform feature extraction preprocessing on the preliminary unstructured data, input the preliminary unstructured data after feature extraction preprocessing into the constructed hidden Markov model, and decode it to obtain the final hidden state sequence, which is the final identified entity; Convert the preliminary unstructured data into a plain text format to obtain plain text data; count all unique terms appearing in the plain text data, map each unique term to an index number, and construct a dictionary; convert the dictionary into a bag-of-words vector representation, where each element of the bag-of-words vector records the number of times the corresponding unique term appears; Define the number of potential topics R; use the bag-of-words vector as input, and based on the topic model, fit the model parameters of the topic model through gibbs sampling or variational inference algorithm. The model parameters include the word distribution of each potential topic and the topic distribution of each preliminary unstructured data; For each potential topic, check the N terms with the largest probability values ​​in its word distribution and use them as the focus words of the potential topic; assign a label to each potential topic; For the preliminary unstructured data, check the B potential topics with the largest probability values ​​in the topic distribution; these potential topics are the main topics to which the preliminary unstructured data belongs; combine the focus words of each main topic into a piece of text as the text representation of the main topic; the text representations of all main topics constitute the text data.

5. The artificial intelligence-based smart employment system according to claim 4, characterized in that: The method of constructing the hidden Markov model includes: Collect m groups of text data with entity annotations as training data; extract training features from the training data; the training features include the current word, part of speech, case mode and whether it contains numbers; use the training features as observations of the hidden Markov model; encode the training features; Define a tag sequence based on training data to represent entity types; use the tag sequence as the hidden state sequence of the hidden Markov model; Initialize the parameters of the hidden Markov model, including the initial state probability, state transition probability and emission probability; The initial state probability π(1) represents the probability of the hidden state at time t=1; the state transition probability A(i,j) represents the probability of transitioning from hidden state i to hidden state j at time t; the emission probability B(j,k) represents the probability of observing the observation value o_k when in hidden state j; where k is the index of the observation value; Using the initialized parameters, the hidden Markov model is trained on the training data; and the Baum-Welch algorithm is used to maximize the log-likelihood of the hidden Markov model on the training data; the optimal parameters of the hidden Markov model are obtained, and the parameters at this time are fixed; and the constructed hidden Markov model is obtained.

6. The intelligent employment system based on artificial intelligence according to claim 5 is characterized in that: The method of decoding to obtain the final hidden state sequence includes: The preliminary unstructured data is constructed into a preliminary observation sequence O of length T, and the preliminary observation sequence O is divided into M segments, each of which is Define Q(l,k) to represent the kth optimal path of the lth segment, and P(l,k) to represent the probability corresponding to Q(m,k); for the first segment, use the Viterbi algorithm to solve the K optimal paths Q(1,k) and the corresponding probabilities P(1,k) of the first L observations, where 1≤k≤K; K refers to the number of optimal paths that need to be retained in each segment; For each subsequent segment l, use the Viterbi algorithm to solve the K optimal paths Q(l,k) and corresponding probabilities P(l,k) of the segment observations; For each optimal path Q(l,k), calculate its combined probability with all K optimal paths in the previous segment; the calculation formula for the combined probability is: C(J,k)=P(l-1,J)×P(l,k)×∑α(t)×P′(q(l-1,J)_L→q(l,k)_1|f(l-1,J),f(l,k),o_L); where C(J,k) represents the probability of the Jth optimal path of the l-1th segment and the kth path of the lth segment to form a complete path, P(l-1,J) represents the probability of the Jth optimal path of the l-1th segment, P(l,k) represents the probability of the kth optimal path of the lth segment, and α( t) is the attention weight at time t; f(l,k) is the feature vector of the kth optimal path of the lth segment, f(l-1,J) is the feature vector of the Jth optimal path of the l-1th segment, and o_L is the observation value at the junction of the two segments; q(l-1,J)_L→q(l,k)_1 means the transition from the last state q(l-1,J)_L of the Jth optimal path of the l-1th segment to the first state q(m,k)_1 of the kth optimal path of the lth segment; P′() is the state transition probability; Keep the K combinations of optimal paths with the largest combined probability C(j,k). Among the K optimal paths of the last segment M, select the optimal path Q(M,k) with the largest probability P(M,k) as the final hidden state sequence. Q(M,k) is the kth optimal path representing the Mth segment, and P(M,k) is the probability corresponding to Q(M,k).

7. The artificial intelligence-based smart employment system according to claim 6, characterized in that: The keyword extraction method includes: Convert the job seeker personal data set into text format to obtain a text set; define a window with a size of r, slide the window in the text set, record the co-occurrence relationship of words in the window, represent each unique word as a node, and if two words co-occur in the window, connect them with an edge, and the weight of the edge is the number of their co-occurrences, to obtain a weighted directed graph, recorded as a word co-occurrence graph; In the word co-occurrence graph, calculate the importance score of each node; the word co-occurrence graph G = (V, E), where V is the set of nodes and E is the set of edges; Define the adjacency matrix A. If there is an edge from node a to node b, then the element A[a,b] in the adjacency matrix is ​​1, otherwise it is 0; For each node a, calculate its comprehensive out-degree d(a) = ∑ b w(a,b)×A[a,b]×pos(a,b); where w(a,b) is the weight of the edge from node a to node b, and pos(a,b) is the part-of-speech weight function of nodes a and b; pos(a,b)=match(ps(a),ps(b))×γ+(1-match(ps(a),ps(b)))×ε; where ps(a) is the part-of-speech tag of node a, and ps(b) is the part-of-speech tag of node b; γ and ε are two weight parameters; match(ps(a),ps(b)) is the matching function, which is 1 when ps(a) and ps(b) are the same, otherwise it is 0; Initialize the importance score W(a) of each node a as Where D is the total number of nodes. In each iteration, the importance score W(a) of each node a is updated according to the importance scores of other nodes. Among them, S(e) is the value of the damping coefficient at the e-th iteration, W(b) is the importance score of node b, and d(b) is the comprehensive out-degree of node b; Where S(e-1) is the value of the damping coefficient at the e-1th iteration, θ is a positive learning rate that controls the amplitude of the adjustment, W(a)_(e) is the importance score of node a at the e-th iteration, and W(a)_(e-1) is the importance score of node a at the e-1th iteration; Repeat the iteration until the importance score of each node converges, that is, the score change between two consecutive iterations is less than the preset change threshold; output the importance score of each node at the last iteration, sort the nodes from high to low according to the importance score, and select the words corresponding to the first H nodes as keywords.

8. The intelligent employment system based on artificial intelligence according to claim 7, characterized in that: The professional competency score is calculated as follows: Define a domain set, including n1 domains, each representing a domain; determine a set of keywords X_c = {x_I} for each domain c, where x_I is the Ith keyword of domain c; Construct an n1×m1 domain-keyword matrix ME, where n1 is the number of domains, m1 is the total number of keywords in all domains, and each element ME_cI of the domain-keyword matrix ME represents the importance weight of the I-th keyword in domain c; construct a job seeker keyword vector Y based on the keywords in the job seeker's personal data set, where the element Y_y=1 in Y indicates that the set of keywords in the job seeker's personal data set contains the y-th keyword, otherwise Y_y=0; Multiply the job applicant keyword vector Y with the field-keyword matrix ME to obtain the job applicant-field vector SH, where the element SH_c in SH represents the job applicant's professional ability score in field c.

9. The artificial intelligence-based smart employment system according to claim 8, characterized in that: The method for generating the job seeker occupation recommendation list includes: Construct a Bayesian network model, define network nodes in the Bayesian network model, the network nodes include nodes mapped to the professional ability scores of job seekers in various fields and nodes mapped to occupations; nodes mapped to the professional ability scores of job seekers in various fields are recorded as professional ability score nodes, and nodes mapped to occupations are recorded as occupation recommendation nodes; conditional probability distributions between given network nodes are given; Normalize the job seeker's professional ability scores in n fields, encode the occupations, and map them into discrete state values; For each career recommendation node h, calculate its conditional probability Z(h|S1,S2,...,Sn); where S1,S2,...,Sn are the first professional ability score node to the nth professional ability score node, that is, all professional ability score nodes; Where Z(S1) is the conditional probability distribution of S1, Z(S2) is the conditional probability distribution of S2, and Z(Sn) is the conditional probability distribution of Sn; Z(h,S1,S2,...,Sn) represents the joint probability distribution of the career recommendation node h and all professional ability score nodes S1,S2,...,Sn; For each occupation, the corresponding conditional probability is calculated, and the occupations are sorted in descending order according to the size of the conditional probability. The first N2 occupations are selected to form a recommended occupation list for job seekers.

10. The artificial intelligence-based smart employment system according to claim 9, characterized in that: The secondary screening method includes: Construct a membrane computing model and define the membrane structure of the membrane computing model, including the membrane hierarchy, membrane initialization objects and membrane computing rules; the membrane hierarchy includes the outermost layer, the middle layer and the innermost layer, the outermost layer is the skin membrane, the middle layer is the working membrane, and the innermost layer is the job seeker membrane; Initialize the skin membrane to an empty set, the job membrane to a job seeker's occupation recommendation list, and the job seeker membrane to a job seeker's personal dataset; The membrane calculation rules include sending rules, matching rules, screening rules and optimization rules; the sending rules are to send the keywords and professional ability scores of job seekers from the job seeker membrane to the job membrane; the matching rules are to calculate the comprehensive matching score of each occupation in the job seeker's occupation recommendation list in the job membrane according to the keywords and professional ability scores of the job seeker; The match score is calculated by: For each occupation v in the job seeker's occupation recommendation list, construct a occupation description vector V_v={u_ov}, where u_ov represents the importance value of the oth keyword to occupation v; For job seekers, all their keywords are concatenated at the vector level to obtain the job seeker's comprehensive key vector Y′; For occupation v, calculate the similarity between the comprehensive key vector of job seeker w and the occupation description vector as the keyword matching score Among them, ||Y′|| and ||V_v|| represent the L2 norm of Y′ and V_v respectively; For occupation v, define the ability weights of the job seeker's professional ability scores in n fields respectively; then calculate the ability matching score between job seeker w and occupation v; The keyword matching score and the ability matching score are weighted and summed to obtain the comprehensive matching score between job seeker w and occupation v; The screening rule is to send occupations with comprehensive matching scores greater than or equal to the preset screening threshold from the work membrane to the skin membrane; the optimization rule is to further screen the occupations in the skin membrane; The membrane calculation process is executed according to the defined membrane structure and membrane calculation rules. The membrane calculation process is an iterative process, and the membrane calculation rules are executed once at each time step. The membrane calculation process is continuously executed until the termination condition is met and the membrane calculation process is terminated. When the membrane calculation process is terminated, the object in the skin membrane is the final smart employment plan. The smart employment plan includes recommended occupations and comprehensive matching scores.

Images